论文解读

AI 辅助整理
方法
构建AgroWM模型,先推断生育阶段再生成图像,并与GPT Image 2.5对
发现
水稻阶段匹配率54.6%,高于GPT Image 2.5的34.8%,9至20周仍达46.4%。
机会点
图像预测尚未实地验证,可结合田间时序数据与遥感验证,并拓展到产量和病害预测。

论文摘要

提出农业世界建模任务,由单张照片预测作物数周至数月长相。AgroWM先推断育阶段再生成图像,水稻阶段匹配率54.6%,高于GPT Image 2.5的34.8%,9至20周仍46.4%。属图像预测,未实地验证。

全文

arXiv

CANWORLDMODELSPREDICTRICEGROWTH? TOWARD
AGRICULTURALWORLDMODELS
Wenkai Ji1, Zeteng Li2, Yinkun Guo2, Jie Zhou1,∗, Liang He1, Zhenong Jin2,∗
1School of Computer Science and Technology, East China Normal University, Shanghai
2Peking University
{jzhou, lhe}@cs.ecnu.edu.cn
Github:https://github.com/ECNU-ICALK/ArgoWM
Model:https://huggingface.co/models/ecnu-icalk/ArgoWM
ABSTRACT
World models can imagine how a physical scene evolves over the next few seconds, but growers need
to know how a crop will develop over weeks to months: will this rice plant have headed a month
from now, and how will it look at harvest? We introduceagricultural world modeling, the task of
predicting how the same plant will look at a requested horizon of days to months, from one photo of
it, its cultivar, and its camera view. In preliminary analysis, even state-of-the-art image generators get
this wrong: they often grow the plant too little or give it the color and form of the wrong growth stage.
We argue that they do not work out the plant’s developmental state, which the photo reveals only
indirectly. We therefore propose AgroWM, which reasons about development before it draws growth.
Its reasoning module, PhenoClock, works out the current growth stage, predicts the future one, and
writes a Growth Script with numeric targets for height, plant area, and color; its generation module,
the Rooted Renderer, then edits the photo to follow this script, adding new organs and colors to the
photographed plant. Because pixel scores can prefer an unchanged copy of the photo, we instead
check each forecast’s traits and developmental timing against the plant’s own later photos. On rice,
54.6% of the forecasts of AgroWM are matched to the target stage (Stage R@1), against 34.8% for
GPT Image 2.5, 24.0% for Qwen Image 2.1, and 47.4% for the strongest domain-specific baseline,
the same backbone fine-tuned without Growth Scripts; 9 to 20 weeks ahead, it still reaches 46.4%,
against 35.7% for that baseline. On maize, a crop with a different body plan and stage scale, the same
design again scores best on all four measures.
1 Introduction
World models [1–3] let machines imagine how a physical scene evolves over the next few seconds, as when a glass falls,
strikes a surface, and shatters (Fig. 1a). Questions about growing food instead require forecasts over weeks to months:
how tall will this plant grow, how much will its canopy spread, and when will it flower and fill its grain? For rice, which
feeds more than half of the world’s population [4], the answers otherwise take up to a season to arrive. We ask whether
a world model can give them in advance: from one photo of a plant, its cultivar, and its camera view, predict how the
same plant will look at a requested horizon of days to months (Fig. 1b). We call this taskagricultural world modeling.
Its answer is a photo, which, unlike a single number, can be checked by eye and measured for any visible trait.
The task is harder than it looks, as shown by a preliminary study of 64 rice cases: 1–4 weeks (22 cases), 5–8 weeks (24),
and 9–20 weeks (18) (Sec. 3).First, two state-of-the-art image generators fail at it: Qwen Image 2.1 hardly changes the
plant, although real plants grow about tenfold in projected area over 9 to 20 rounded weeks (median 11.2×), and GPT
Image 2 grows it too little at long horizons, shows senescence-related color changes at the wrong stage, and redraws it
toward a generic rice plant (Fig. 2).Second, the horizon alone does not decide the future stage: of the 18 cases with
horizons of 9 to 20 weeks, 6 plants have not yet headed by the target date, while 9 are already filling grain or ripening. A
∗Corresponding authors.
arXiv:2611.26145v1 [cs.CV] 19 Nov 2026

ICALK@ECNU
(a) Conventional world model (b) Agricultural world model
now +0.1 s +0.2 s
seconds:physical dynamics
today +7 weeks +9weeks +12weeks
weeks to months: the plant grows driven by a growth script
Figure 1.From physical dynamics to growth.(a) A conventional world model predicts rapid physical events. (b) From one photo,
an agricultural world model predicts how the plant will look at each requested horizon of weeks to months, as it moves through
growth stages.
forecaster must therefore infer each plant’s developmental state, which the photo reveals only indirectly.Third, common
image scores reward doing nothing: a copy of today’s photo has the highest SSIM in Tab. 1, and Qwen Image 2.1,
which barely changes the plant, the highest PSNR.
A reliable growth forecast must get the individual plant, its future traits, and their timing right. Existing methods
address only parts of this problem. Time-lapse generators [5, 6] draw plausible growth from text, with no particular
plant or date to match, and keyframe interpolation [ 7] and 4D reconstruction [ 8] are given the later images that a
forecast has to predict. Plant-growth GANs [9–11] come closest, but they map elapsed time, treatment, or simulated
biomass to pixels without reasoning about development. Agricrafter [12] generates controllable videos of a crop’s full
growth cycle, but a coherent sequence alone does not show whether a given plant reaches the correct stage within the
requested horizon. Crop models [13, 14] do track development using temperature and day length, but require cultivar-
and site-specific parameters and do not directly generate the future appearance of an individual plant. Our approach
follows an agronomist’s reasoning by first predicting the plant’s future developmental stage and then generating an
image consistent with that prediction. For example, if a plant is currently tillering and is expected to reach jointing in
four weeks, the forecast image should show elongated stems and a taller plant.
We turn this insight into AgroWM, an agricultural world model with two modules (Fig. 3). First, its reasoning module,
PhenoClock, is a vision-language model that works out the plant’s current stage, predicts its stage at the horizon, and
writes this chain of thought (CoT) as a Growth Script. The script is a short, structured note: the current and future
stage and, for each visible trait such as height, plant area, and color, today’s value, the target value, and the change
in words. Since no dataset records such notes, PhenoClock learns from scripts written in hindsight: for each training
pair, the stages and values come from records and measurements of the real future, and a language model adds the
words (Sec. 4.2). Second, its generation module, the Rooted Renderer, then edits today’s photo as the script says; it is
prompted and trained to keep the plant’s form, framing, and scale, on pairs chosen to reward real growth (Sec. 4.3).
To judge forecasts, we check what matters to a grower: the same plant, at the right size and color, at the right stage of
development (Sec. 3.2). Timing is checked against the plant’s own photo series: Stage R@1 is the share of forecasts
whose most similar photo of the plant lies in the target stage. On held-out plants of the public rice dataset WGSR [15],
AgroWM reaches 54.6% Stage R@1, against 34.8% for GPT Image 2.5 and 24.0% for Qwen Image 2.1, two state-of-
the-art generators, and 47.4% for the strongest domain-specific model, the same backbone fine-tuned on the same pairs
without Growth Scripts (Tab. 1). Reasoning about development thus adds 7.2 points (95% CI 5.2–9.2), and 10.7 points
9 to 20 weeks ahead (Fig. 4). On maize, which grows a single stalk with a tassel instead of tillers and panicles, the same
design again scores best, so the approach is not tied to one crop (Sec. 5.3).
In summary, we make three contributions:
• We explore agricultural world modeling and find that the current image generators do not pace a plant’s growth to the
horizon, reach its target stage, or keep the individual.
• We propose AgroWM, which reasons before it renders: PhenoClock forecasts the developmental stage with the plant’s
height, area, and color as an agricultural CoT learned from scripts written in hindsight, and the Rooted Renderer turns
this forecast into an edited photo of the plant; it gives the best forecasts on both rice and maize.
• We show that standard metrics such as PSNR and SSIM are useless in guiding the generation of ecologically
meaningful plants, and introduce an evaluation of identity, geometry, color, and timing against the plant’s growth
series.
2

ICALK@ECNU
2 Related Work
2.1 World Models in Computer Vision
World models and long-horizon prediction.World models learn how an environment evolves, first as latent
simulators for control [ 16, 17] and now as large video generators presented as world simulators [ 1, 18, 19] for
games [20, 21], driving [2], and robots [3, 22]. Longer horizons have been reached by temporal abstraction, which
predicts directly at coarse or adaptive steps [23–27], and learned weather models forecast physical fields days ahead [28,
29]. These models primarily predict physical dynamics, including object motion, interactions, and the evolution of
physical fields.We open a regime they do not cover: the growth of an organism over days to months, driven by a state
that the image shows only indirectly.
Probing and evaluating world knowledge.Video generators often violate physical laws despite realistic frames [ 30–
32] and lack a consistent time scale [33]. Time-lapse generation [5, 6, 34, 35] targets slow processes from text, with no
individual’s real future as ground truth, keyframe interpolation [7] is given the last frame, and face aging [36] involves
no new organs. These tasks often use pixel and perceptual metrics [37–39] to measure visual similarity, but such metrics
can favor blurry averages under uncertainty [40, 41]. In addition, FID [42] compares image distributions and reflects
both visual fidelity and diversity, but does not separate their contributions to the score. We extend this probing from
physical consistency to the timing of biological development. Genotype, environment, and management shape a plant’s
developmental trajectory, so elapsed time alone does not determine its future stage. A forecaster must infer where the
plant is in its development and how far it will progress over the requested horizon. A realistic image can still be the
wrong forecast: its size, color, or organs may be plausible but mistimed. We therefore score timing by retrieval in the
same plant’s photo series (Sec. 3.2).
Reasoning before generation.Multimodal language models have been used to plan image edits [ 43, 44] and to
reason before synthesis [45], extending chain-of-thought prompting [46], and CausalWM [47] reasons about causal
variables before predicting video. Their reasoning concerns layout, semantics, or short-term physics.PhenoClock
differs in what it reasons about and where its supervision comes from: it advances a hidden developmental state by a
requested horizon and states it in image measurements, in a schema designed with agronomic knowledge. Its training
scripts are written backwards from real future images around recorded stages and measured values, in the spirit of
privileged information [48, 49] and rationalization [50].
2.2 Modeling Growth in Agriculture
Process-based crop and plant models.Agronomists have long forecast crops with process-based models: crop
models such as ORYZA2000 [13], DSSAT [14], APSIM [51], and WOFOST [52] advance a developmental state with
temperature and day length and simulate biomass and yield from weather, soil, and structural plant models simulate
architecture [53, 54]. They need cultivar- and site-specific parameters and output numbers or geometry rather than the
appearance of a particular plant.We keep their central idea, a developmental state advanced through time, but infer it
from the photo and render it as an image.
Observing and forecasting growth from images.Image-based phenotyping describes the plant’s current state:
vision models segment organs, measure traits, and estimate plant age from photographs [55–57]. Growth forecasting
goes further by predicting its later appearance. Plant-growth GANs predict future images from earlier photos, using
fixed prediction intervals or explicit time, treatment, and biomass conditions, and assess their outputs through image
quality and trait agreement [9–11]. Agricrafter [12] generates full-growth-cycle crop videos with diffusion models, and
4D reconstruction [8] recovers growth that has already been imaged. We explicitly forecast the future developmental
state of the plant to guide image generation from a current photo at horizons of up to 20 weeks, and evaluate the
developmental timing of the plant’s photo series, on two crops.
3 Preliminary Analysis
We first check how state-of-the-art image generators forecast growth and whether standard scores reflect their mistakes.
We test two of them: GPT Image 2, the general-purpose model with the highest Stage R@1 in Tab. 1, and Qwen Image
2.1, the base of our Rooted Renderer. The 64 validation cases come from 64 cultivars, half in each view, with 22, 24,
and 18 cases at 1–4, 5–8, and 9–20 rounded weeks (the longest is 15 weeks; Fig. 2). Each receives today’s photo and a
prompt with the cultivar, view, and horizon, plus a keep-list that asks to keep the plant’s identity and to avoid premature
yellowing or panicles.
3

ICALK@ECNU
Today Real future Qwen Image 2.1 GPT Image 2 AgroWM Evidence on all 64 cases
Growth does not keep up with the horizon① → AgroWM: advance the stage by Δ
a +8 weeks · jointing → grain filling · top view Pixel scores rank the only forecast at the right size and stage last
area ×1
area ×6.8
×0.7
no change✗
×2.8
too little✗
×6.6
Score Qwen GPT AgroWM
PSNR ↑ 11.4 11.1 10.0
SSIM ↑ 0.692 0.649 0.530
Area err. ↓ 20.4 13.3 0.5
Extent err. ↓ 40.2 16.5 0.0
Yellow err. ↓ 17.4 37.6 2.4
Stage rank ↓ 4/8 4/8 1/8
Event rank ↓ 5/8 5/8 1/8
Forecasts often show the wrong growth stage② → AgroWM: stage first, then traits
b +8 weeks · tillering → jointing · front view GPT draws panicles too early; color and both ranks catch it, SSIM does not
yellow 11%
yellow 16%
46%
too little✗
31%
heads early✗
21%
Score Qwen GPT AgroWM
PSNR ↑ 12.2 12.2 12.6
SSIM ↑ 0.756 0.688 0.615
Area err. ↓ 16.5 8.4 8.1
Extent err. ↓ 30.5 12.1 4.6
Yellow err. ↓ 29.8 14.5 4.7
Stage rank ↓ 5/17 12/17 3/17
Event rank ↓ 5/17 12/17 3/17
Forecasts often show a different plant③ → AgroWM: keep the plant, add organs
c +3 weeks · tillering → jointing · top view Qwen and GPT win the pixel scores but miss size, color, and stage
area ×1
area ×6.5
×0.9
no change✗
×2.3
template✗
×6.9
Score Qwen GPT AgroWM
PSNR ↑ 18.2 17.9 15.5
SSIM ↑ 0.901 0.913 0.828
Area err. ↓ 5.0 3.7 0.3
Extent err. ↓ 46.0 9.7 8.8
Yellow err. ↓ 20.4 11.9 0.4
Stage rank ↓ 2/9 2/9 1/9
Event rank ↓ 4/9 4/9 1/9
Best per row: Qwen GPT AgroWM
1–4 wk 5–8 wk 9–20 wk
×2.7
×4.4
×11.2
×2.4 ×3.0
×5.6
×0.9 ×0.9 ×1.0
Visible-area growth (median)①
Real future GPT Image 2 Qwen Image 2.1
vegetative, front view grain filling, ripening
16
54
25 2422 23
Yellow/brown share (%)②
Real future GPT Image 2 Qwen Image 2.1
How often does no growth win?④
% of the 64 cases in which Qwen beats AgroWM
PSNR 58
SSIM 97
DINO sim. 6
Area err. 19
Extent err. 23
Yellow err. 28
late stages 18
Stage R@1 5
Event R@1 0
→ AgroWM: score size and developmental timing
Figure 2.How existing models fail to forecast growth, and why standard scores miss it.
3.1 How Existing Models Fail
① Growth does not keep up with the horizon.The further ahead a forecast looks, the more the plant should change,
but neither generator keeps pace with the real change. Qwen Image 2.1 barely changes the plant: its median projected
area is 0.93× to 0.98× today’s, while real plants grow 2.7×, 4.4×, and 11.2× in the three ranges. GPT Image 2 grows
the plant with the horizon but falls further behind, reaching2.4×,3.0×, and5.6×(Fig. 2a).
② Forecasts often show the wrong growth stage.How a plant looks depends on its growth stage, and the generators
often draw the wrong one. Although the prompt forbids it, GPT Image 2 yellows young plants early: in 8 of the 15
front-view cases that stay vegetative, the share of yellow or brown plant pixels in its forecast exceeds the real plant’s by
more than 10 percentage points (Fig. 2b). Both models keep ripening plants green: in the 22 cases at grain filling or
ripening, the real plants are 54% yellow or brown on average, against 24% for GPT Image 2 and 23% for Qwen Image
2.1 (② in Fig. 2). The horizon alone cannot fix the stage: at 9–20 rounded weeks, 6 of the 18 plants have not yet headed,
while 9 are filling grain or ripening, so each plant follows its own clock.
③Forecasts often show a different plant.A forecast must show the same individual plant, but GPT Image 2 often
changes its form toward a typical rice plant. In front views it draws a taller, upright plant, more than 10 points of image
height too tall in 13 of the 32 cases (Fig. S1e); in top views it draws a symmetric rosette (Fig. 2a, c). Qwen Image 2.1
keeps the individual mostly by not changing it.
Together, these failures suggest that the generators jump from today’s photo and the horizon straight to a new photo,
without working out the plant’s developmental state, and fall back on a generic picture of rice. Our method is therefore
built around three requirements: (T1) advance the developmental state explicitly (①); (T2) read the plant’s own stage
first and derive its traits from it (②); and (T3) keep the individual and change only what development changes (③).
3.2 Our Evaluation Design
Assessing whether forecasts meet these requirements calls for more than visual similarity. We first examine where
standard scores fall short, then define an evaluation of identity, size, color, and developmental timing.
④ Standard scores reward no growth.Standard scores do not catch these failures: pixel scores favor an unchanged
photo, and the other scores each see only part of the change. In the 64 cases, plants cover only 14% of the real future
4

ICALK@ECNU
4.1 Task Definition
Test-time input
xt
cultivar a
view v
horizon Δ+49 d
x̂t+Δ= f(xt, Δ, c)
Training pairs
xt xt+Δ
stages z, slots m
4.2 Reasoning Module: PhenoClock ①②
Forecasting from today's photo
PhenoClock g
vision-language model
s = g(xt, Δ, c)
vision
LLM ŝ
Lclock: next-token loss on s*
Scripts written in hindsight
Hindsight Tutor h
language model
43%
89% z: S2 → S4
Δ: 49 days
m: slots
words w Π+ z, m s*
leakage control · verification
✓change, organs, color
✗numbers, dates, counts, poses
↻validator rejects, Tutor retries
Growth Script s
ŝteststrain
Measured chain of thought
Stage locate · advance
S2 jointing → S4 heading
Growth plan observe · quantify
slots in image units
Culm 43 → 89%
Leaves/area2.2 → 11.0%
Density 22.6 → 21.7%
Panicle 0 → 1
Color 14.7 → 26.7%
Culm: “…culms elongate upward
from the same basal crown …”
Target phenotype describe
Panicle: “Branched panicles are
visibly emerged …”
4.3 Generation Module: Rooted Renderer ①③
Keeping the same plant
Taska, v, Δ
Keep-listsame plant, framing;
no reverse growth
Growth Script schanges
text encoder
xt
edit image,
always given
VAE
Diffusion transformer vθ
Forecast
x̂t+Δ
x̂t+Δ= r(xt, Δ, c, s)
Learning to follow the script
flow matching to xt+Δgiven s
; at test time ŝtakes its place
Pairs that reward real growth
40%
1–4 wk
35%
5–8 wk
25%
9–20 wk
MILP: each stage as often
a target as a source
✗≤ 3 days apart
✗backward stage
✗early yellowing
✗height loss
64% 48%
✗removed
selected pairs
test time training only
trained
frozen S2, 43%recorded stages and measured values
Figure 3.AgroWM forecasts development before rendering growth.PhenoClock writes a Growth Script with the future stage and
targets for height, plant area, and color, and the Rooted Renderer edits the source photo to follow it. Example: a validation pair 49
days apart (7 weeks in the prompt), with the Hindsight Tutor’s reference script (center) and the forecast from PhenoClock’s own
script (right).
photo, so PSNR and SSIM mostly measure the white background, and an unchanged copy of today’s photo gets the
highest mean PSNR and SSIM of all forecasts we score (14.9 dB and 0.81, against 14.2 dB and 0.75 for GPT Image
2). DINOv2 similarity [39] depends less on the background, but its scale shifts with the horizon: the copy scores 0.74
at 1–4 rounded weeks and 0.41 at 9–20, so a value means little without its horizon. Single traits see growth but not
development: GPT Image 2 beats the copy and Qwen Image 2.1 in projected area in 52 of the 64 cases, yet in 33 of
them the forecast looks most like a photo from another stage (Fig. 2b).
What must a forecast get right?We score four parts of a forecast separately.(i)Identity( ③): DINOv2 feature
similarity (DINO) and foreground-mask IoU with the real future.(ii)Size and shape( ①, ③): errors, in points of
the image, of projected area, plant height, and density, the plant’s share of its convex hull.(iii)Color( ②): error of
the yellow–brown share of plant pixels.(iv)Timing( ①, ②): we rank the photos of the same plant and view, today’s
included (14 on average), by DINOv2 ViT-B/14 similarity to the forecast; Stage R@1 and Event R@1 are the shares of
forecasts whose top photo lies in the target stage or is the real target. Since these photos share the plant and camera, a
forecast that does not grow matches today’s photo: Qwen Image 2.1 reaches an Event R@1 of only 1.1%, worse than
picking a photo at random (at least 7%; Tab. 1). On the 2,000 validation queries, PSNR and SSIM prefer this no-growth
forecast to AgroWM on 51% and 93%, the timing scores on 6% and 1% (Tab. S1; ④ in Fig. 2), so PSNR and SSIM are
for reference only. A Growth Script states the target size and color, so these errors may favor AgroWM; the timing
scores use only the forecast image and recorded labels.
4 Method
AgroWM answers two questions in turn, what the plant will become and how it will then look (Fig. 3), instead of
asking one generator to guess the developmental state, the amount of growth, and what to keep all at once (Sec. 3). The
reasoning module, PhenoClock (Sec. 4.2), answers the first: from today’s photo and the query, it predicts a measured
Growth Script (T1, T2). The generation module, the Rooted Renderer (Sec. 4.3), answers the second: it edits the same
photo as the script says (T3). In training, a Hindsight Tutor that sees both photos of a pair writes the words of its
reference script around recorded stages and measured values.
4.1 Task Definition
The model sees one photo and predicts the plant at a requested future horizon. Given a photo xt of a plant at time t, a
horizon ∆, and conditions c= (a, v), the cultivar a and camera view v∈ {front,top}, an agricultural world model
predicts the same plant at timet+ ∆:
ˆxt+∆ =f(x t,∆, c),1≤∆≤T,(1)
5

ICALK@ECNU
where T is the rest of the season. In the rice experiments, elapsed intervals are rounded to whole weeks for the model
input, while target photos retain their recorded observation dates. The model must infer the current stage from the
photo, since no stage label is given. The recorded stages and slot values (Tab. S2) of a pair (xt, xt+∆) are available
only in training and for evaluation. AgroWM splitsfinto its two modules,
s=g(x t,∆, c),ˆx t+∆ =r(x t,∆, c, s),(2)
where PhenoClockgwrites the Growth Scriptsand the Rooted Rendererredits the source photo accordingly.
4.2 Reasoning Module: PhenoClock
PhenoClock answers the first question, what the plant will become. It is a vision-language model whose output, a
structured chain of thought, spells out how the plant will develop. Unlike free-form reasoning before generation [45],
this chain of thought must be exact enough to draw and to check, and WGSR does not provide it. We therefore define
what it contains, write training examples in hindsight, and teach PhenoClock to produce them from today’s photo.
Measured chain of thought.Like an agronomist’s note, a Growth Script states the stage first and the traits second, in
three blocks of fixed order,
s=

zt, zt+∆;{(d k, mkt , mkt+∆)}k;{e k}k,(3)
where k runs over the slots Kv of view v: culm height, plant area, density or canopy, panicles, and color (Tab. S2).
The stage block gives today’s and the future stage on eight stages S1–S8 adapted from the IRRI scale [58] (tillering,
jointing, booting, heading, grain filling, milk and dough, yellow ripening, and maturity); the growth-plan block gives,
for each slot, a sentence dk on the visible change with today’s and the target value; and the target-phenotype block
describes the end state (ek). Read in order, a Growth Scriptlocatesthe plant,advancesit by ∆, and thenobserves,
quantifies, anddescribeseach slot. All values except the panicle label, which follows the cultivar’s heading date, are
image measurements, such as the plant’s extent from the bottom edge and its share of the image: each is a target for
PhenoClock, a condition for the Rooted Renderer, and a quantity measured again on the forecast (Sec. 3.2).
Scripts written in hindsight.We write the training scripts backwards, from pairs whose future we know. A language
model explains well but measures poorly, so the Hindsight Tutor h sees both photos and the facts of a pair, the recorded
stages z= (z t, zt+∆), the interval, and the slot values m= (m t, mt+∆), but writes only the words w, in a strict
schema; a fixed templateΠthen inserts the stages and values and computes each change:
w=h

xt, xt+∆,∆, z, m

, s ⋆ = Π(w, z, m).(4)
The stages and numbers of s⋆ are thus copied from the records and measurements. Becauseh sees the answer, its prompt
allows words only on what a forecaster could know: the kind and direction of change, organ emergence consistent
with the panicle value, and color progression, never numbers, dates, organ counts, or future poses (leakage control). A
validator checks the structure, the stages against the records, and forbidden forms such as numbers in the words, and h
rewrites rejected scripts (verification).
Forecasting from today’s photo.At test time the future is unknown, so PhenoClock must write the same Growth
Script from today’s photo, the horizon, and the conditions alone. We fine-tune PhenoClock with a next-token loss on s⋆
whose tokens fall into three groups, the stage names, the metric fields, and the remaining script text:
Lclock = 0.1L stage + 0.4L metric + 0.5L text,(5)
where each term is the loss−logp g

s⋆i |s ⋆<i, xt,∆, c over the tokenss⋆i of one group. The future reaches PhenoClock
only through this label, as in learning with privileged information [48, 49], and the label’s source-side stage and values
also teach PhenoClock to read today’s plant.
4.3 Generation Module: Rooted Renderer
The Rooted Renderer answers the second question, how the plant will then look: it should add the organs and colors
that the Growth Script states and keep the rest of today’s photo, whereas the generators of Sec. 3 either leave the plant
unchanged (①) or alter it (③). We build it from a pretrained image generation and editing model with three choices, one
per paragraph below: it edits today’s photo, it learns from pairs that reward real growth, and it trains on the Hindsight
Tutor’s scripts.
6

ICALK@ECNU
Table 1.Main comparison on rice.Bold: best per column; PSNR and SSIM for reference only.
Identity Geometry Color Timing Pixel (ref.)
Method DINO↑IoU↑Area↓Height↓Density↓Yellow↓ StageR@1↑ EventR@1↑ PSNR↑SSIM↑
General-purpose
GPT Image 2 0.709 0.303 5.59 11.49 6.31 24.40 35.3 12.8 13.23 0.689
GPT Image 2.5 0.721 0.299 6.92 12.57 6.70 29.17 34.8 10.8 12.92 0.694
Qwen Image 2.1 0.693 0.241 7.85 18.28 6.83 32.31 24.0 1.114.200.748
Qwen-Image-Edit-2511 0.568 0.188 28.84 20.46 25.08 33.76 26.5 9.2 9.54 0.476
FLUX.2 Klein 0.496 0.259 41.58 22.47 36.51 34.11 25.2 9.6 7.36 0.392
FLUX 3 Image Klein Base 0.683 0.244 13.87 17.65 12.38 35.14 24.1 1.4 12.98 0.691
LANCE 0.636 0.240 23.81 20.90 24.17 31.61 24.5 3.8 10.55 0.591
Domain-specific
Copy 0.702 0.252 7.17 19.06 6.24 29.33 23.2 0.0 14.020.750
LANCE FT 0.8150.3423.91 8.04 5.62 15.95 46.9 16.9 13.94 0.709
FLUX.2 Klein FT 0.767 0.305 4.70 11.00 6.03 24.59 39.3 18.0 13.91 0.697
Qwen Image 2.1 FT 0.834 0.323 3.73 7.83 5.33 11.43 47.4 22.5 14.10 0.685
Agricrafter 0.666 0.329 5.20 9.61 7.97 15.79 32.1 13.3 14.11 0.706
Ours
AgroWM0.8600.3273.51 7.43 4.98 10.00 54.6 24.814.14 0.686
Keeping the same plant.The Rooted Renderer edits today’s photo instead of drawing a new one: the source photo is
the edit image in training and at test time, and every target keeps the frame of its source, in which the numbers of the
Growth Script are defined. The prompt states the task (accession, view, and horizon in whole weeks), a fixed keep-list
(same identity, background, framing, and scale; development only forward, without premature yellowing or panicles),
and the Growth Script. A prompt alone is not enough, since GPT Image 2 received the same keep-list and still altered
the plant (Sec. 3); the Rooted Renderer instead learns to follow it from its training pairs.
Pairs that reward real growth.A generator learns whatever its training pairs reward. Among the candidate pairs,
however, long horizons are scarce, short ones are nearly solved by copying, and early stages appear mostly as sources
and late ones as targets. We therefore fix a target share for each horizon range and, within each split, select candidate
pairs within weekly quotas by a mixed-integer program that makes each stage appear about as often as a source as a
target, with at least one pair per plant (Eq. (7) in Sec. D). From the candidates, we drop pairs whose change is not
growth, such as backward stages or early yellowing, and draw the final pairs stratified by source and target stage; the
fine-tuned baselines use them too.
Learning to follow the script.The Rooted Renderer should learn that following the script pays off, so we train it with
the Hindsight Tutor’s scripts s⋆, whose stages and values match the target, and give it the scripts ˆsof PhenoClock only
at test time; imperfect predicted scripts in training could teach it to ignore the script. We fit each triplet (xt, s⋆, xt+∆)
with the flow-matching objective of the base model [59, 60]:
Lren =E τ,ϵ

vθξτ , τ;E(xt), p⋆ −(ϵ−ξ 0)

22,(6)
where E is its V AE encoder,ξ0 =E(x t+∆), ξτ = (1−τ)ξ 0 +τ ϵwith ϵ∼ N(0, I), and p⋆ is the prompt built from
(∆, c, s⋆); the text and V AE encoders stay frozen.
5 Experiments
We compare AgroWM with state-of-the-art image generators and domain-specific models on rice and maize, test it at
long horizons, and assess its modules and the accuracy of its Growth Scripts.
5.1 Experimental Setup
Datasets.We validate AgroWM on two crops that grow very differently, with two public datasets of the Open Plant
Image Archive (OPIA) [61]. WGSR [15] photographs 93 rice cultivars about once a week over the season; we keep its
top view and one side (front) view with white backgrounds, labeled with the eight growth stages of Sec. 4.2 and each
cultivar’s heading date. WGSM [62] photographs 224 maize inbred lines across nine growth stages, from vegetative
7

ICALK@ECNU
growth to physiological maturity, and we keep one side view. Each query pairs an earlier and a later photo of the same
plant and view; rice pairs are 4 to 140 days apart (median 35). We split by plant: rice has 20,000 training queries from
252 plants and 2,000 validation queries from 79 held-out plants, 78 of whose cultivars also occur in training, and maize
has 6,400 and 1,600 queries from 348 and 84 plants (Sec. D).
Baselines.The baselines fall into the two groups of Tab. 1.General-purposegenerators are state-of-the-art image
generation and editing models: the closed-source GPT Image 2 and GPT Image 2.5 and the open-source Qwen Image
2.1, Qwen-Image-Edit-2511 [63], FLUX.2 Klein, FLUX 3 Image Klein Base [64], and LANCE [65]. Each receives
today’s photo and a prompt with the cultivar, view, and horizon, plus the Rooted Renderer’s keep-list, without a Growth
Script (Sec. E).Domain-specificmethods are built for this task. Copy returns today’s photo. LANCE FT, FLUX.2 Klein
FT, Qwen Image 2.1 FT, and Agricrafter [12], a crop growth video generator, are fine-tuned on our data; method-specific
training details are given in Sec. E.
Implementation details.We fine-tune PhenoClock from Qwen3.8-27B [ 66] with LoRA adapters [67] and the Rooted
Renderer from Qwen Image 2.1 [ 63, 68], a generation and editing model with a 7B diffusion transformer, in all
transformer weights; the Hindsight Tutor is GPT-6.1 Sol [69]. Both modules are trained for two epochs on the 20,000
training queries with a global batch of 16. At test time, the Rooted Renderer samples 20 denoising steps; the settings of
all methods are in Sec. E.
5.2 Main Results
AgroWM outperforms all seven state-of-the-art generators on every identity, geometry, color, and timing score (Tab. 1),
with Stage and Event R@1 of 54.6% and 24.8%, against 34.8% and 10.8% for GPT Image 2.5 and 24.0% and 1.1% for
Qwen Image 2.1. These generators change the plant too little or without control: Qwen Image 2.1 barely changes it, the
other open models miss its area by 14 to 42 points, two to six times the error of the copy, and GPT Image 2, the best of
them in Stage R@1, reaches 35.3%. Fine-tuning on our pairs helps all three generators we fine-tune, raising Stage R@1
from 24.5% to 46.9% for LANCE, from 25.2% to 39.3% for FLUX.2 Klein, and from 24.0% to 47.4% for Qwen Image
2.1, all above the 32.1% of Agricrafter. Adding Growth Scripts on top of the same backbone and pairs lifts AgroWM
above every domain-specific model, by at least 7.2 and 2.3 points in Stage and Event R@1. It also has the best identity
and trait scores of all generated forecasts, except IoU, where LANCE FT and Agricrafter are slightly higher (0.342 and
0.329 against 0.327).
5.3 Generalization to Maize
Table 2.Generalization to maize.
Method DINO↑Area↓Height↓Stage R@1↑
Copy 0.522 3.62 33.02 0.0
GPT Image 2 0.727 2.63 16.64 21.1
Qwen Image 2.1 FT 0.764 1.97 11.12 30.1
AgroWM0.795 1.73 10.61 30.9
Maize tests the framework on a crop with a single stalk,
tassels, and ears. We train maize-specific models from the
original pretrained backbones, adapting the stage scale and
organ descriptions (Secs. D and E). AgroWM has the best
overall means on the four reported metrics (Tab. 2): Stage
R@1 is 30.9%, against 30.1% for Qwen Image 2.1 FT and
21.1% for GPT Image 2, with area and height errors of
1.73 and 10.61 points. Its Stage R@1 advantage over Qwen
Image 2.1 FT is smaller than on rice.
5.4 Long-Horizon Prediction
The further ahead a forecast looks, the less of the answer today’s photo shows, so Fig. 4 splits the scores by horizon.
One to four weeks ahead, the stage often stays the same, so the copy of today’s photo already reaches 45.8% Stage
R@1, and AgroWM and Qwen Image 2.1 FT are about equally accurate (Stage R@1 62.8% and 62.4%, Event R@1
27.3% and 27.8%).
8

ICALK@ECNU
AgroWM Qwen Image 2.1 FT GPT Image 2 Copy
Stage R@1 (%) ↑
0
20
40
60
1–4 wk 5–8 wk 9–20 wk
+10.7
Event R@1 (%) ↑
0
10
20
30
1–4 wk 5–8 wk 9–20 wk
+4.8
Area error (pp) ↓
0
5
10
15
1–4 wk 5–8 wk 9–20 wk
−0.4
Yellow error (pp) ↓
0
25
50
1–4 wk 5–8 wk 9–20 wk
−1.9
Figure 4.Forecast quality by horizon on rice.
As the horizon grows, all four methods of Fig. 4 lose Stage
R@1, AgroWM the least: from 1–4 to 9–20 weeks, it falls
from 62.8% to 46.4%, against 62.4% to 35.7% for Qwen
Image 2.1 FT and 49.2% to 22.4% for GPT Image 2. From
five weeks on, AgroWM is best on all four scores of Fig. 4.
Its Stage R@1 lead over Qwen Image 2.1 FT is 0.4 points
at 1–4 weeks but 13.9 and 10.7 points at 5–8 and 9–20
weeks, and at 9–20 weeks it also leads by 4.8 points in Event
R@1 and lowers the area and yellow errors by 0.4 and 1.9
points. Since the two share backbone, pairs, and schedule
and differ in whether they use Growth Scripts, reasoning
about development helps most where today’s photo leaves
the future open.
5.5 Ablation Study
Table 3.Ablation studies on rice.R@1 in %, area error in
points.
Variant StageR@1↑ EventR@1↑ Area↓DINO↑
AgroWM (full) 54.6 24.8 3.51 0.860
w/o PhenoClock 47.4 22.5 3.73 0.834
w/o Rooted Renderer 24.8 12.3 8.22 0.565
Hindsight Tutor (Upper Bound) 58.9 28.6 2.92 0.874
We compare AgroWM with a script-free fine-tuning control,
a script-only base generator, and a target-informed script
reference in Tab. 3. Without PhenoClock (Qwen Image
2.1 FT, trained and run without Growth Scripts), Stage and
Event R@1 drop by 7.2 and 2.3 points (95% CI 5.2–9.2 and
0.5–4.2, paired bootstrap over plants), and Stage R@1 by
10.7 points at 9–20 weeks (Sec. 5.4). Without the Rooted
Renderer, the base Qwen Image 2.1 receives the Growth
Scripts of PhenoClock in its prompt but neither the source
photo nor fine-tuning: Stage and Event R@1 fall to 24.8%
and 12.3%, the area error rises to 8.22 points, and DINO
drops to 0.565, so a good script is not enough without a
renderer that keeps the plant. With the Hindsight Tutor’s target-informed scripts, the same Rooted Renderer reaches
58.9% and 28.6% (Sec. 5.6).
5.6 How Accurate Is the Growth Script?
Table 4.Growth Script accuracy on the rice validation
split, against the recorded stages and measured slots of the
real future (height: 1,000 front views).±1: within one stage.
Similarity: mean cosine similarity of BGE [70] embeddings to
the Hindsight Tutor’s 18,000 slot descriptions.
Current Future Future±1
Stage accuracy (%)↑89.90 78.90 99.40
Height Area Yellow
Target MAE (pp)↓4.81 2.16 8.24
Change-direction Macro-F1↑0.659 0.708 0.556
Semantic similarity (cosine)↑0.842
Errors in the Growth Script carry over to the forecast, so we
check the scripts of PhenoClock against the recorded stages
and measured traits of the real future (Tab. 4). PhenoClock
names today’s stage correctly in 89.9% of the queries and
the future stage in 78.9%, and 99.4% of its future stages
are at most one stage off, so advancing the stage is harder
than reading it. Its target values are off by 4.81, 2.16, and
8.24 points for height, area, and yellow, and 1,999 of its
2,000 scripts pass the strict format check. These errors
cost 4.3 points of Stage R@1: with the Rooted Renderer
fixed, the Hindsight Tutor’s scripts reach 58.9% and those
of PhenoClock 54.6% (95% CI of the gap 2.2–6.3; Tab. 3),
so better scripts would give better forecasts.
6 Conclusion and Future Work
We asked whether world models can forecast the slow change of living things, and found that two state-of-the-art
image generators either barely change a rice plant or change it at the wrong pace and stage, while pixel scores favor the
unchanged plant. AgroWM narrows this gap by reasoning before rendering: it first forecasts a Growth Script of the
plant’s development, learned from scripts written in hindsight, and then edits the photo to match. Reasoning adds 7.2
points of Stage R@1 over the same backbone without Growth Scripts, most beyond four weeks, and puts AgroWM
ahead of state-of-the-art generators and domain-specific models on both rice and maize, two crops with different body
plans and stage scales. Because each forecast is a photo of the same plant, we can check its identity, size, color, and
developmental timing against the plant’s own later photos. The rice results come from potted plants on one greenhouse
9

ICALK@ECNU
platform, with the background removed, mostly of cultivars seen in training, and from pairs filtered for growth. Future
work will condition forecasts on weather and management records and test them on new cultivars, more species, and
field images.
References
[1] Tim Brooks, Bill Peebles, et al. Video generation models as world simulators. OpenAI technical report, 2024.
[2] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and
Gianluca Corrado. GAIA-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080,
2023.
[3] Sherry Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and
Pieter Abbeel. Learning interactive real-world simulators. InICLR, 2024.
[4] Sumithra Muthayya, Jonathan D. Sugimoto, Scott Montgomery, and Glen F. Maberly. An overview of global rice
production, supply, trade, and consumption.Annals of the New York Academy of Sciences, 1324(1):7–14, 2014.
[5] Shenghai Yuan, Jinfa Huang, Yujun Shi, et al. MagicTime: Time-lapse video generation models as metamorphic
simulators.IEEE TPAMI, 2025.
[6] Shenghai Yuan, Jinfa Huang, Yongqi Xu, et al. ChronoMagic-Bench: A benchmark for metamorphic evaluation
of text-to-time-lapse video generation. InNeurIPS Datasets and Benchmarks Track, 2024.
[7] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M.
Seitz. Generative inbetweening: Adapting image-to-video models for keyframe interpolation. InICLR, 2025.
[8] Weihan Luo, Lily Goli, Sherwin Bahmani, Felix Taubner, Andrea Tagliasacchi, and David B. Lindell. Grow with
the flow: 4D reconstruction of growing plants with Gaussian flow fields.arXiv preprint arXiv:2602.08958, 2026.
[9] Lukas Drees, Laura Verena Junker-Frohn, Jana Kierdorf, and Ribana Roscher. Temporal prediction and evaluation
of Brassica growth in the field using conditional generative adversarial networks.arXiv preprint arXiv:2105.07789,
2021.
[10] Lukas Drees, Dereje T. Demie, Madhuri R. Paul, Johannes Leonhardt, Sabine J. Seidel, Thomas F. Döring, and
Ribana Roscher. Data-driven crop growth simulation on time-varying generated images using multi-conditional
generative adversarial networks.Plant Methods, 20:93, 2024.
[11] Xinyi Wang et al. GAN-based image prediction of maize growth across varieties and developmental stages.Plant
Methods, 21:110, 2025.
[12] Lining Wang, Zhan Zhang, and Hongxun Yao. Agricrafter: Full-growth-cycle crop video generation via diffusion
models.Computers and Electronics in Agriculture, 239:110946, 2025.
[13] B. A. M. Bouman, M. J. Kropff, T. P. Tuong, M. C. S. Wopereis, H. F. M. ten Berge, and H. H. van Laar.
ORYZA2000: Modeling Lowland Rice. International Rice Research Institute, Los Baños, Philippines, 2001.
[14] James W. Jones, Gerrit Hoogenboom, Cheryl H. Porter, et al. The DSSAT cropping system model.European
Journal of Agronomy, 18(3–4):235–265, 2003.
[15] Zhixin Tang, Zhuo Chen, Yuan Gao, Ruxian Xue, Zedong Geng, Qingyun Bu, Yanyan Wang, Xiaoqian Chen,
Yuqiang Jiang, Fan Chen, Wanneng Yang, and Weijuan Hu. A strategy for the acquisition and analysis of
image-based phenome in rice during the whole growth period.Plant Phenomics, 5:0058, 2023.
[16] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. InNeurIPS, 2018.
[17] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world
models.arXiv preprint arXiv:2301.04104, 2023.
[18] NVIDIA. Cosmos world foundation model platform for physical AI.arXiv preprint arXiv:2501.03575, 2025.
[19] Wan Team. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,
2025.
[20] Jake Bruce, Michael Dennis, Ashley Edwards, et al. Genie: Generative interactive environments. InICML, 2024.
[21] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines.
arXiv preprint arXiv:2408.14837, 2024.
[22] Mido Assran et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning.arXiv
preprint arXiv:2506.09985, 2025.
[23] Alexander Neitz, Giambattista Parascandolo, Stefan Bauer, and Bernhard Schölkopf. Adaptive skip intervals:
Temporal abstraction for recurrent dynamical models. InNeurIPS, 2018.
10

ICALK@ECNU
[24] Karol Gregor, George Papamakarios, Frederic Besse, Lars Buesing, and Théophane Weber. Temporal difference
variational auto-encoder. InICLR, 2019.
[25] Dinesh Jayaraman, Frederik Ebert, Alexei A. Efros, and Sergey Levine. Time-agnostic prediction: Predicting
predictable video frames. InICLR, 2019.
[26] Vaibhav Saxena, Jimmy Ba, and Danijar Hafner. Clockwork variational autoencoders. InNeurIPS, 2021.
[27] Jesse Farebrother, Matteo Pirotta, Andrea Tirinzoni, Marc G. Bellemare, Alessandro Lazaric, and Ahmed Touati.
Compositional planning with jumpy world models.arXiv preprint arXiv:2602.19634, 2026.
[28] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, et al. Learning skillful medium-range global weather
forecasting.Science, 382(6677):1416–1421, 2023.
[29] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global
weather forecasting with 3D neural networks.Nature, 619(7970):533–538, 2023.
[30] Hritik Bansal, Zongyu Lin, Tianyi Xie, et al. VideoPhy: Evaluating physical commonsense for video generation.
arXiv preprint arXiv:2406.03520, 2024.
[31] Bingyi Kang, Yang Yue, Rui Lu, et al. How far is video generation from world model: A physical law perspective.
arXiv preprint arXiv:2411.02385, 2024.
[32] Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models
understand physical principles? InWACV, 2026.
[33] Mingyang Wu, Siyuan Yang, Jiongze Yu, Pardis Taghavi, Fangzhou Lin, and Zhengzhong Tu. The pulse of
motion: Measuring physical frame rate from visual dynamics.arXiv preprint arXiv:2603.14375, 2026.
[34] Yipin Zhou and Tamara L. Berg. Learning temporal transformations from time-lapse videos. InECCV, 2016.
[35] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using
multi-stage dynamic generative adversarial networks. InCVPR, 2018.
[36] Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, and Ira Kemelmacher-Shlizerman. Lifespan age
transformation synthesis. InECCV, 2020.
[37] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error
visibility to structural similarity.IEEE TIP, 13(4):600–612, 2004.
[38] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness
of deep features as a perceptual metric. InCVPR, 2018.
[39] Maxime Oquab, Timothée Darcet, Théo Moutakanni, et al. DINOv2: Learning robust visual features without
supervision.Transactions on Machine Learning Research, 2024.
[40] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. InCVPR, 2018.
[41] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. InICML, 2018.
[42] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by
a two time-scale update rule converge to a local Nash equilibrium. InNeurIPS, 2017.
[43] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based
image editing via multimodal large language models. InICLR, 2024.
[44] Yuzhou Huang, Liangbin Xie, Xintao Wang, et al. SmartEdit: Exploring complex instruction-based image editing
with multimodal large language models. InCVPR, 2024.
[45] Rongyao Fang, Chengqi Duan, Kun Wang, Linjiang Huang, Hao Li, Shilin Yan, Hao Tian, Xingyu Zeng, Rui
Zhao, Jifeng Dai, Xihui Liu, and Hongsheng Li. GoT: Unleashing reasoning capability of multimodal large
language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025.
[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V . Le, and
Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. InNeurIPS, 2022.
[47] Ziming Xu, Shuang Liang, Ruobing Han, Ziqiao Xi, Mingxing Rao, Kun Zhou, Zijun Zhang, Yuchen Yan, Yufan
Wei, Junbo Huang, Yifei Shao, Fang Nan, and Biwei Huang. CausalWM: Causal chain-of-thought reasoning for
embodied world model.arXiv preprint arXiv:2609.23184, 2026.
[48] Vladimir Vapnik and Akshay Vashist. A new learning paradigm: Learning using privileged information.Neural
Networks, 22(5–6):544–557, 2009.
[49] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krähenbühl. Learning by cheating. InCoRL, 2019.
11

ICALK@ECNU
[50] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. STaR: Bootstrapping reasoning with reasoning. In
NeurIPS, 2022.
[51] Dean P. Holzworth, Neil I. Huth, Peter G. deV oil, et al. APSIM – evolution towards a new generation of agricultural
systems simulation.Environmental Modelling & Software, 62:327–350, 2014.
[52] Allard de Wit, Hendrik Boogaard, Davide Fumagalli, et al. 25 years of the WOFOST cropping systems model.
Agricultural Systems, 168:154–167, 2019.
[53] Przemyslaw Prusinkiewicz and Aristid Lindenmayer.The Algorithmic Beauty of Plants. Springer, New York,
1990.
[54] J. V os, J. B. Evers, G. H. Buck-Sorlin, B. Andrieu, M. Chelle, and P. H. B. de Visser. Functional–structural plant
modelling: A new versatile tool in crop science.Journal of Experimental Botany, 61(8):2101–2115, 2010.
[55] Hanno Scharr, Massimo Minervini, Andrew P. French, et al. Leaf segmentation in plant phenotyping: A collation
study.Machine Vision and Applications, 27(4):585–606, 2016.
[56] Wanneng Yang, Zilong Guo, Chenglong Huang, et al. Combining high-throughput phenotyping and genome-wide
association studies to reveal natural genetic variation in rice.Nature Communications, 5:5087, 2014.
[57] Ruchi Bhatt, Shreya Bansal, Amanpreet Chander, Rupinder Kaur, Malya Singh, Mohan Kankanhalli, Abdulmo-
taleb El Saddik, and Mukesh Kumar Saini. GroMo: Plant growth modeling with multiview images.arXiv preprint
arXiv:2503.06608, 2025.
[58] International Rice Research Institute, Manila, Philippines.Standard Evaluation System for Rice (SES), 5th edition,
2013.
[59] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative
modeling. InICLR, 2023.
[60] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data
with rectified flow. InICLR, 2023.
[61] Yongrong Cao, Dongmei Tian, Zhixin Tang, Xiaonan Liu, Weijuan Hu, Zhang Zhang, and Shuhui Song. OPIA:
An open archive of plant images and related phenotypic traits.Nucleic Acids Research, 52(D1):D1530–D1537,
2024.
[62] Weixuan Wang, Weijun Guo, Liang Le, Jia Yu, Yue Wu, Dongwei Li, Yifan Wang, Huan Wang, Xiaoduo Lu, Hong
Qiao, Xiaofeng Gu, Jian Tian, Chunyi Zhang, and Li Pu. Integration of high-throughput phenotyping, GW AS, and
predictive models reveals the genetic architecture of plant height in maize.Molecular Plant, 16(2):354–373, 2023.
[63] Chenfei Wu, Jiahao Li, Jingren Zhou, et al. Qwen-Image technical report.arXiv preprint arXiv:2508.02324, 2025.
[64] Black Forest Labs. FLUX 3 Image Klein Base. Hugging Face model card, 2026.
[65] Fengyi Fu, Mengqi Huang, Shaojin Wu, Yunsheng Jiang, Yufei Huo, Hao Li, Yinghang Song, Fei Ding, Jianzhu
Guo, Qian He, Zheren Fu, Zhendong Mao, and Yongdong Zhang. Lance: Unified multimodal modeling by
multi-task synergy.arXiv preprint arXiv:2605.18678, 2026.
[66] Qwen Team. Qwen3.8-27B. Hugging Face model card, 2026.
[67] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu
Chen. LoRA: Low-rank adaptation of large language models. InICLR, 2022.
[68] Qwen Team. Qwen-Image-2.1. GitHub repository, 2026.
[69] OpenAI. Introducing GPT-6.1 Sol. OpenAI blog, 2026.
[70] Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. C-Pack: Packed
resources for general Chinese embeddings. InProceedings of the 47th International ACM SIGIR Conference on
Research and Development in Information Retrieval, 2024.
12

ICALK@ECNU
Supplementary Material
Today Real future Qwen Image 2.1 GPT Image 2 Today Real future Qwen Image 2.1 GPT Image 2 Evidence on all 64 cases
Growth does not keep up with the horizon①
→ AgroWM: advance the stage by Δ
(a) +3 weeks: tillering → jointing (front view)
area ×1
area ×2.6
no change✗
×0.9
too much✗
×4.8
(b) +11 weeks: tillering → booting (top view)
area ×1
area ×65
no change✗
×1.5
too little✗
×6.0
Forecasts often show the wrong growth stage②
→ AgroWM: stage first, then traits
(c) +11 weeks: tillering → jointing (front view)
yellow 25%
yellow 22%
no change✗
27%
already ripe✗
72%
(d) +12 weeks: jointing → ripening (top view)
yellow 23%
yellow 93%
stays green✗
32%
stays green✗
19%
Forecasts often show a different plant③
→ AgroWM: keep the plant, add organs
(e) +4 weeks: heading → grain filling (front view)
height 58%
height 58%
57%
new plant✗
87%
(f) +10 weeks: tillering → heading (top view)
area ×1
area ×11
no change✗
×1.0
template✗
×8.2
1–4 wk 5–8 wk 9–20 wk
×2.7
×4.4
×11.2
×2.4 ×3.0
×5.6
×0.9 ×0.9 ×1.0
Visible-area growth (median)①
Real future
GPT Image 2
Qwen Image 2.1
vegetative, front view grain filling, ripening
16
54
25 2422 23
Yellow/brown share (%)②
Real future
GPT Image 2
Qwen Image 2.1
SSIM ↑ PSNR ↑ Size error ↓
63
43
12
1
21
52
Evaluation: cases won (of 64)④
No growth (copy, Qwen)
GPT Image 2
→ AgroWM: score size and developmental timing
Figure S1.How existing models fail to forecast growth, two single cases per failure, front and top views.Red frames mark
failures, and the images of a case share one crop. Right: statistics over all 64 cases; ④ counts the cases in which each score ranks no
growth or GPT Image 2 best.
A More Cases of the Preliminary Analysis
Figure 2 in Sec. 3 shows one case per failure, chosen so that AgroWM is best on every trait error and retrieval rank
shown. In its top-view rows (a) and (c), PSNR and SSIM also prefer Qwen Image 2.1 or GPT Image 2, and this holds in
only 3 of the 64 cases; in its front-view row (b), one of only 2 of the 32 front views in which every plant and timing
score prefers AgroWM, SSIM prefers Qwen Image 2.1 while PSNR prefers AgroWM. Figure S1 shows two cases per
failure, chosen for the failure alone and covering both views, including the front-view failures that Sec. 3.1 reports as
statistics. Figure S2 shows all three, rows (a) and (c) of Fig. 2 and one front view, with the score labels in full and the
share of cases in which each score prefers no growth, grouped as in Tab. S1.
The cases of Fig. S1 add details that Sec. 3.1 gives only as statistics. GPT Image 2 paces its edits only loosely by the
interval: after three weeks, it adds panicles to a plant that is still jointing (Fig. S1a), whereas after eleven weeks, it keeps
a young tillering plant at 6× its area, against 65× for the real plant (Fig. S1b). In front views, it turns a plant that is still
jointing 72% yellow or brown, against 22% in reality (Fig. S1c), and draws a plant taller, reaching 87% of the image
height where the real plant stays at 58% (Fig. S1e). Ripening plants stay green in both models (Fig. S1d), most of all in
top views, where the real plants are on average 63% yellow or brown, against 9% for GPT Image 2. Qwen Image 2.1
keeps the plant mostly by not changing it, but in 4 of the 64 cases it replaces the plant with a single culm or a field
texture.
B Analysis of Each Score
This section expands Sec. 3.2. We compare two forecasts with little or no growth, the exact copy of today’s photo and
Qwen Image 2.1, which barely changes it, with that of AgroWM, which sees only today’s photo, the horizon, and the
conditions: Tab. S1 counts how often each score prefers no growth over the 2,000 validation queries, and the 64 cases
below show why. All scores are computed by one evaluator on the white-background images.
13

ICALK@ECNU
Today Real future Qwen Image 2.1 GPT Image 2 AgroWM
a +8 weeks · jointing → grain filling · top view Pixel scores rank the only forecast at the right size and stage last
Score Qwen GPT AgroWM
PSNR ↑ 11.4 11.1 10.0
SSIM ↑ 0.692 0.649 0.530
Area err. ↓ 20.4 13.3 0.5
Extent err. ↓ 40.2 16.5 0.0
Yellow err. ↓ 17.4 37.6 2.4
Stage rank ↓ 4/8 4/8 1/8
Event rank ↓ 5/8 5/8 1/8
b +2 weeks · tillering → jointing · front view Both baselines grow the plant too little, yet win both pixel scores
Score Qwen GPT AgroWM
PSNR ↑ 20.7 20.3 19.1
SSIM ↑ 0.935 0.936 0.891
Area err. ↓ 2.0 1.9 0.5
Extent err. ↓ 5.1 10.3 3.9
Yellow err. ↓ 37.6 4.2 1.3
Stage rank ↓ 1/10 1/10 1/10
Event rank ↓ 2/10 1/10 1/10
c +3 weeks · tillering → jointing · top view Qwen and GPT win the pixel scores but miss size, color, and stage
Score Qwen GPT AgroWM
PSNR ↑ 18.2 17.9 15.5
SSIM ↑ 0.901 0.913 0.828
Area err. ↓ 5.0 3.7 0.3
Extent err. ↓ 46.0 9.7 8.8
Yellow err. ↓ 20.4 11.9 0.4
Stage rank ↓ 2/9 2/9 1/9
Event rank ↓ 4/9 4/9 1/9
Best per row: Qwen GPT AgroWM
How often does no growth win?
Share of the 64 cases in which Qwen Image 2.1,
which barely changes the plant, beats AgroWM
(lower is better) 0 50% 100%
Standard
PSNR 58
SSIM 97
DINO sim. 6
Plant traits (ours)
Area err. 19
Extent err. 23
Yellow err. 28
late stages 18
Timing (ours)
Stage R@1 5
Event R@1 0
Dashed: 50%. Late stages: the 22 cases that reach
grain filling or ripening. DINO and timing: 63 cases.
Figure S2.What each score sees.The three cases in which PSNR and SSIM prefer Qwen Image 2.1 or GPT Image 2 while
AgroWM is best or tied on every plant and timing score; (a) and (c) are rows (a) and (c) of Fig. 2.
Table S1.Which scores reward no growth?Left: share of the
validation queries on which a score strictly prefers the copy or
Qwen Image 2.1 to AgroWM; ties count in the denominator but
not as wins. Right: validation means (errors in percentage points,
R@1 in %). Height on 1,000 front views, DINO and R@1 on
1,964 queries, the rest on 2,000.
No growth wins (%) Validation mean
Score Copy Qwen Qwen AgroWMStd.
PSNR↑45.5 50.8 14.20 14.14
SSIM↑89.9 93.1 0.748 0.686
DINO sim.↑25.7 18.7 0.693 0.860Traits
Area err.↓31.3 25.6 7.85 3.51
Height err.↓29.6 29.1 18.28 7.43
Density err.↓45.0 41.2 6.83 4.98
Yellow err.↓28.8 21.3 32.31 10.00Time
Stage R@1↑5.4 5.6 24.0 54.6
Event R@1↑0.0 0.6 1.1 24.8
Background agreement can dominate pixel scores.
Over the 64 cases, the copy has a mean PSNR of 14.9 dB
and a mean SSIM of 0.81, against 14.7 dB and 0.78 for
Qwen Image 2.1 and 14.2 dB and 0.75 for GPT Image
2. Plants cover on average 14% of a real future photo
in the 64 cases, and 17% over the validation set; the rest
is white background. PSNR and SSIM therefore weigh
mostly pixels that a forecast leaves untouched when it
does not grow, and they penalize a leaf drawn slightly off
twice, where it is and where it should be, but a missing
leaf only once. SSIM prefers the copy to AgroWM in 63
of the 64 cases, and PSNR in 38. In Fig. S2a, the only
forecast with the right size and stage receives the lowest
PSNR and SSIM of the three, 1.3 dB below Qwen Image
2.1. Over the 2,000 validation queries, Qwen Image 2.1,
which barely changes the plant, has the highest PSNR
of all deployable models in Tab. 1, only the copy has a
higher SSIM, and both are above AgroWM.
Global features are not anchored in time.Cosine
similarity of DINOv2 features [39] between the forecast
and the real future depends little on the background and
prefers the copy to AgroWM in only 8 of 63 cases. Its value, however, depends on the horizon as much as on the
forecast: the copy scores 0.74 at 1–4 weeks and 0.41 at 9–20 weeks, so the same value can mean a good long-horizon
forecast or no change at a short horizon. Over the validation set, the copy reaches 0.702, higher than five of the seven
general-purpose generators and below only GPT Image 2 and GPT Image 2.5 (0.709 and 0.721). A single similarity
also cannot say in which direction a forecast is wrong, whether too early, too late, or not the same plant.
Plant traits measure growth, but no single trait measures time.Measured on the plant pixels alone, the errors in
visible area, vertical extent, and yellow–brown share see the growth that pixel scores miss: in Fig. S2a Qwen Image 2.1
is 20.4 and 40.2 points off in area and extent, against 0.5 and 0.0 for AgroWM. Each trait, however, sees only one side
of development. A forecast of the right size can be at the wrong stage: GPT Image 2 has a smaller area error than the
copy and Qwen Image 2.1 in 52 of the 64 cases (5.4 points on average over the 64 cases, against 10.0 for the copy), yet
33 of these 52 forecasts miss the target stage under Stage R@1. Size cannot tell a green plant from a ripening one: in a
top-view case twelve weeks ahead, GPT Image 2 misses the real area by only 3.6 points, but its plant is green where the
14

ICALK@ECNU
Table S2.Slots of a Growth Script, image proxies measured on the plant pixels of the white-background photo (F: front, T: top
view). Bottom: one growth-plan line of a Hindsight Tutor script, validation pair, tillering to jointing in 27 days.
Slot View Definition Unit
Culm F plant height (bottom edge to top plant pixel) % of image height
Leaves/area F, T all plant pixels / image pixels % of image
Plant density F plant pixels / pixels of their convex hull %
Canopy T plant pixels / pixels of their convex hull %
Panicle F, T on or after the accession’s heading date 0/1
Color development F, T yellow and brown / plant pixels, HSV thresholds %
Culm: Culms extend upward from the same basal crown as the established tillers progress from short basal growth toward jointing-
stage elongation. Metric = Height; Source = 23.44%; Target = 62.39%; Change = +38.95 pp.
real one is 94% yellow and brown. Color, in turn, is informative only when color changes: over all 64 cases Qwen
Image 2.1 beats AgroWM on yellow–brown error in 18, because most real futures are still green, but in only 4 of the 22
cases that reach grain filling or ripening, such as Fig. S2a, where it is 17.4 points off and AgroWM 2.4. In front views,
every plant reaches the bottom edge (all 64 source and target images), so the vertical extent equals the culm slot of the
Growth Script there.
Timing scores and no growth.The 36 of 2,000 queries whose plant has fewer than five photos in the view are
skipped. Since the gallery contains today’s photo, the copy has an Event R@1 of zero, and its Stage R@1 is one only
where the stage does not change, which is then the right answer. Accordingly, Event R@1 never prefers the copy to
AgroWM, and Stage R@1 does so in 3 of 63 cases, all with an unchanged stage.
C Slots of a Growth Script
Table S2 lists the slots of a Growth Script (Sec. 4.2) and how each is measured.
Table 4 in Sec. 5.6 scores the Growth Scripts of PhenoClock against the recorded stages and measured slots of the real
future. Most of its future-stage errors are off by one stage, its direction of change is most often right for area and least
for yellow, its descriptions read much like the Hindsight Tutor’s, and the one script that fails the strict format check
states a height change that does not match its values and is passed to the Rooted Renderer as written.
Change-direction Macro-F1.We classify each predicted and reference source-to-target change as decrease ( d <
−1 pp), stable (−1≤d≤1 pp), or increase (d >1pp), where d is the change in the corresponding image proxy. For
each of height, area, and yellow–brown share, Macro-F1 is the unweighted mean of the F1 scores of these three classes.
Description similarity.We align the predicted and Hindsight Tutor descriptions by view, section (Growth plan or
Target phenotype), and slot, and remove the numeric evidence suffixes before comparison. The 2,000 validation queries
contain 18,000 reference descriptions across these two sections: five slots per front view and four per top view. We
encode the descriptions with BAAI/bge-small-en-v1.5 [70], use L2-normalized CLS embeddings, and report
the mean cosine similarity over matched descriptions. This score measures agreement with the Hindsight Tutor’s
wording and meaning; the stage and numeric scores assess factual accuracy.
D Datasets, Cleaning, and Training Pairs
WGSR [15], a public dataset of the Open Plant Image Archive [61], holds 16,758 RGB photos of 93 rice accessions,
taken from two sides and from the top about once a week from 42 to 182 days after sowing, on a LemnaTec phenotyping
platform in a greenhouse in Beijing. We keep the top view and one side view, called the front view, and generate
white-background versions with Qwen Image 2.1 [68], whose prompt, rewritten per photo by its prompt enhancer, asks
to keep the plant unchanged and to replace everything else with white, so that methods and scores see only the plant.
Some outputs keep remnants of the pot or rails or shift the color of young leaves; the color-coverage gate and the pair
gates below remove part of them. Each photo carries one of the eight growth stages of Sec. 4.2, and each accession
its heading date. To test on unseen plants, one random eligible plant per accession goes to validation before pairs are
filtered and selected, and all photos of a plant stay in one split. Stage-stratified sampling yields 10,000 training and
1,000 validation pairs, each in both views: 20,000 training queries from 252 plants of 85 accessions and 2,000 validation
queries from 79 plants, one from each of 79 accessions, 78 of which also occur in training. These queries use 7,540
photos of 331 plants from 86 accessions.
15

ICALK@ECNU
Before pairing, we drop the photos recorded as abnormal growth (three accessions and 57 single photos), all photos
from the day a plant first received an artificial support, and a few others by hand. Within each split, the candidate
selection of Sec. 4.3 solves
min
y∈{0,1}P
P
z
N srcz −N tgtz +λ Pz |N srcz − ¯Nz|+|N tgtz − ¯Nz|,(7)
subject to weekly quotas and at least one pair per plant, where y selects among P candidates, N srcz and N tgtz count
selected sources and targets in stage z, ¯Nz is the count under the pool’s stage proportions, and λ weighs the second
term. It targets 40%, 35%, and 25% of the pairs at 1–4, 5–8, and 9–20 weeks; after the gates and the stratified sampling,
46%, 33%, and 21% of the validation queries fall in these ranges, so 54.4% are beyond four weeks. The gates drop
pairs whose change is not growth: photos at most three days apart; backward stages or panicles; early yellowing, that
is, a yellow–brown share of at least 0.35 in tillering or jointing, or a rise of at least 0.20 from these stages within two
stages; a front-height loss of at least 5 points toward a target before yellow ripening; and color coverage below 70%.
We attribute such changes to background removal, lighting, or leaning plants rather than growth; a pair that kept them
would teach the Rooted Renderer to draw the artifact and the Hindsight Tutor to explain it.
WGSM [62], also available through OPIA, contains images of 224 maize inbred lines acquired on a high-throughput
phenotyping platform across nine growth stages, which cover vegetative growth, tasseling, silking, grain filling, and
physiological maturity. We retain one side view and construct strictly forward pairs within each plant; the 768×1,270
source images are padded by 10 pixels at the top to a 768×1,280 canvas, then processed to a white background with
connected-component cleanup. The resulting queries cover 432 plants, split by plant into 348 training and 84 held-out
validation plants, with 6,400 and 1,600 queries (Sec. 5.3). The protocol maps stages S1–S9 to V3, V5, V7, V10, V14,
VT, R1, R3, and R6, respectively, and uses the unrounded interval in days; the Planner receives the plant ID, source
image, and interval, but not the recorded query stages or future image.
E Implementation Details
For rice, photos are 672×896 pixels (width×height) in the front view and 896×672 in the top view. The candidate
selection of Eq. (7) uses λ= 0.02 ; the Hindsight Tutor runs with reasoning effort xhigh, and all 22,000 queries end
with a Hindsight Tutor script that passes the validator. Table S3 summarizes the training recipes for the learned methods
reported in the paper.
At test time, the Rooted Renderer and Qwen Image 2.1 FT sample 20 denoising steps with a guidance scale of 1 in
bf16, at the source size. Among the open general-purpose baselines, Qwen Image 2.1, Qwen-Image-Edit-2511, and the
9B FLUX.2 Klein base model use 20 denoising steps in bf16, with true classifier-free guidance 1.0, true guidance 4.0,
and guidance 4.0, respectively; Qwen Image 2.1 outputs are composited on white. The 3B LANCE uses its official
image-editing pipeline with 20 steps. GPT Image 2 is called through its image-editing API with automatic quality, and
its outputs are resized to the source size with Lanczos filtering. GPT Image 2.5 and FLUX 3 Image Klein Base are
included as evaluated in the main table; settings not present in the run records are not inferred here. The rice image
generators receive the keep-list of the Rooted Renderer’s prompt and the horizon in rounded whole weeks. All forecasts
are scored at the source size. For FLUX 3 Image Klein Base, only 998 queries have a valid height and 1,996 a valid
density and yellow share.
F Baseline Experiments
This section describes the baselines reported in the paper’s main comparison and maize comparison (Tab. 2). The
copy returns the source image unchanged. The general-purpose image generators receive the source image and the
task prompt described in Sec. 5.1, including the accession, view, horizon, and the Rooted Renderer keep-list, without
a predicted Growth Script. The open-model inference settings and output-size handling are given above; generated
images are evaluated at the source dimensions.
The rice Qwen Image 2.1 FT, FLUX.2 Klein FT, and LANCE FT baselines are trained on the same rice training split
without a Planner or Growth Script; their training recipes are listed in Tab. S3. Qwen Image 2.1 FT is the Simple
Executor control for the Rooted Renderer. The GPT Image API baselines and the copy do not use these training updates.
Agricrafter is a video-generation baseline adapted to WGSR, rather than an image-pair fine-tune; its window construction
and training recipe are listed in Tab. S3. At inference it predicts nine frames with DDIM for 50 steps and recursively
advances by at most eight observation steps; the target frame is selected by the recorded observation-index difference,
not by a requested number of days or weeks.
16

ICALK@ECNU
Table S3.Training configurations for paper methods.All local Planner and Executor runs use bfloat16. The image-model
schedules use a peak learning rate followed by the listed warmup, stable, and cosine-decay fractions.
Model / arm Initialization and trainable
components
Training data Recipe
Rice Planner Qwen3.8-27B; LoRA
r= 32,α= 64; all linear
layers; vision tower and
aligner frozen
20,000 queries; global
batch 16
2 epochs, 2,500 updates; LR10 −4, 3% warmup, cosine
to10 −6; grouped loss0.1/0.4/0.5
Rice Simple
Executor
(Qwen Image
2.1 FT)
Original Qwen Image 2.1;
full DiT; text and V AE
encoders frozen; no Growth
Script
20,000 queries; global
batch 16
2 epochs, 2,500 updates; peak LR10 −5; 10% warmup,
60% stable, 30% cosine to10 −6
Rice CoT
Executor
(Rooted
Renderer)
Original Qwen Image 2.1;
full DiT; text and V AE
encoders frozen; teacher
Growth Scripts in training
20,000 queries; global
batch 16
2 epochs, 2,500 updates; peak LR10 −5; 10% warmup,
60% stable, 30% cosine to10 −6; 8 GPUs
Rice FLUX.2
Klein FT /
LANCE FT
Corresponding pretrained
image model; fine-tuned
image-generation backbone;
no Planner or Growth Script
20,000 queries; global
batch 16
2 epochs, 2,500 updates; LR10 −5; 10% warmup, 60%
stable, 30% cosine to10 −6; 8 GPUs
Agricrafter DynamiCrafter_512
initialization; UNet and
image projector trained
2,108
nine-observation
windows from 178
rice training plants;
global batch 4
6,000 updates; white letterbox canvas320×512
Maize Planner Qwen3.8-27B; LoRA
r= 32,α= 64; all linear
layers; vision tower and
aligner frozen
6,400 queries; global
batch 16
2 epochs, 800 updates; LR10 −4, 3% warmup, cosine to
10−6; grouped loss0.1/0.4/0.5
Maize Simple /
CoT Executors
Separately initialized Qwen
Image 2.1 full DiTs; text and
V AE encoders frozen; no
script / teacher script in
training
6,400 queries per arm;
global batch 16
2 epochs, 800 updates; peak LR10 −5; 10% warmup,
60% stable, 30% cosine to10 −6
The maize comparison contains the source copy, GPT Image 2, a Simple Qwen Image 2.1 Executor, and AgroWM. The
Planner, Simple Executor, and CoT Executor are each trained separately from their original pretrained model on the
maize split; no rice-fine-tuned checkpoint is transferred. Their configurations are listed in Tab. S3. The CoT Executor is
trained on teacher scripts but receives only the source-only Planner prediction at inference; GPT Image 2 uses the same
source image and Simple task prompt, while the copy is unchanged.
G Planner Output Templates
Each response contains one <GROWTH_COT> block with three sections in order:Stage reasoning,Growth plan, and
Target phenotype. Figure S3 shows the crop- and view-specific templates with their ordered fields and value syntax;
current/future stages and numeric targets are predictions, not labels supplied with the query.
For rice, the Planner receives the source image, accession, view, and rounded interval in weeks. Growth plan fields
may include an optionalGuardrails:clause after the change description; unavailable metric evidence is omitted. Rice
numeric proxies use two decimal places, while Panicle usesvisible/not visibleand a signed visibility change. The
top-view template uses Canopy density and omits the front-view Culm/Height field.
For maize, the query supplies the source image, plant ID, side view, and exact interval in days. The four numeric
fields use one decimal place; Tassel and Ear contain only qualitative visibility descriptions, and hidden organs are not
reconstructed. When the Culm change is within 2% of the source proxy, its numeric triplet may be replaced by the
exact wordingStable within the registered two-percent proxy tolerance; Source/Target/Change suppressed.The maize
Planner’s system prompt includes a general guide listing all nine stage IDs S1–S9, their corresponding growth stages,
and visible descriptions; it does not identify the current or future stage of the query plant. The dataset protocol maps
S1–S9 to V3, V5, V7, V10, V14, VT, R1, R3, and R6 at nominal 25, 32, 50, 59, 70, 78, 85, 91, and 113 days after
sowing; these protocol reference days do not imply that every plant was imaged on those dates.
17

ICALK@ECNU
H Evaluation Coverage
The table reports the valid query count for each score in the paper; unavailable phenotype measurements remain missing
and are not counted as zero error. Rice has 2,000 validation queries in total. Height applies to front views only, while
DINOv2 and stage/event retrieval require an eligible same-plant, same-view photo series.
Table S4.Evaluation coverage in the paper.Counts are valid queries over the indicated dataset.
Dataset / method Valid queries by score
Rice, most methodsN= 2,000 : area, density, yellow errors; IoU, PSNR, SSIM; N= 1,000 : height (front); N= 1,964 :
DINOv2, Stage R@1, Event R@1
Rice, FLUX 3N= 998: height;N= 1,996: density and yellow
Maize, all methodsN= 1,600: DINOv2, Stage R@1, pixel scores
Maize, copy / GPT Image 2N= 1,600: area, culm-proxy height, density, yellow errors
Maize, Simple / AgroWMN= 1,596each: area, culm-proxy height, density;N= 1,592/1,591: yellow error
Maize phenotype counts are lower for trained image models when the foreground is empty or does not contain enough
valid color pixels; the underlying predictions remain in the evaluation set. The maize evaluation uses one view, and its
gallery has one image per plant, view, and stage, so Stage and Event retrieval hits coincide; the paper reports Stage
R@1.
18

ICALK@ECNU
(a) Rice / front view
Inputs: image, accession, view, weeks
<GROWTH_COT>
Stage reasoning
Current stage: {current_stage}.
Future stage: {future_stage}.
Growth plan
Target phenotype
</GROWTH_COT>
(b) Rice / top view
Inputs: image, accession, view, weeks
<GROWTH_COT>
Stage reasoning
Current stage: {current_stage}.
Future stage: {future_stage}.
Growth plan
Target phenotype
</GROWTH_COT>
(c) Maize / side view (sv1)
Inputs: image, plant ID, sv1, days
<GROWTH_COT>
Stage reasoning
Current stage: {current_stage}.
Future stage: {future_stage}.
Growth plan
Target phenotype
</GROWTH_COT>
Culm: {visible culm change}.
Metric = Height;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Leaves/area: {leaf area change}.
Metric = Foreground area;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Plant density: {projection change}.
Metric = Solidity;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Panicle: {visibility change}.
Metric = Panicle visibility;
Source = {v_s}; Target = {v_t};
Change = {d_v}.
Plant color development:
{visible color change}.
Metric = Yellow-brown fraction;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Leaves/area: {leaf area change}.
Metric = Foreground area;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Canopy: {projected canopy change}.
Metric = Canopy density;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Panicle: {visibility change}.
Metric = Panicle visibility;
Source = {v_s}; Target = {v_t};
Change = {d_v}.
Plant color development:
{visible color change}.
Metric = Yellow-brown fraction;
Source = {s}%; Target = {t}%;
Change = {d} pp.
Culm: {visible culm change}.
Source={s}%; Target={t}%;
Change={d} pp.
Visible foreground area:
{visible foliage change}.
Source={s}%; Target={t}%;
Change={d} pp.
Projected fill: {projection change}.
Source={s}%; Target={t}%;
Change={d} pp.
Tassel: {source visibility}.
{target visibility}.
Ear: {source visibility}.
{target visibility}.
Visible yellow-brown composition:
{visible color change}.
Source={s}%; Target={t}%;
Change={d} pp.
Culm: {visible endpoint}.
Leaves/area: {visible endpoint}.
Plant density: {visible endpoint}.
Panicle: {visible endpoint}.
Plant color development:
{visible endpoint}.
Leaves/area: {visible endpoint}.
Canopy: {visible endpoint}.
Panicle: {visible endpoint}.
Plant color development:
{visible endpoint}.
Culm: {visible endpoint}.
Visible foreground area: {endpoint}.
Projected fill: {visible endpoint}.
Tassel: {visible endpoint}.
Ear: {visible endpoint}.
Visible yellow-brown composition:
{visible endpoint}.
Braces denote generated text or values. s / t: source / target proxy; d: change in percentage points (pp).
Rice: 2 decimals, signed d; v_s / v_t: visible or not visible, d_v: −1 / +0 / +1. Maize: 1 decimal; stages S1–S9.
Figure S3.CoT output templates for rice and maize.Each column shows the complete three-section wrapper and ordered fields.
Braces denote generated descriptions or values rather than literal output tokens; each field occupies one logical line and wraps here
for readability. Growth plan describes the transition and gives proxy evidence, while Target phenotype describes the visible endpoint
qualitatively. Image-proxy definitions are in Tab. S2.
19

论文原文与出版信息

arXiv
查看论文原文

继续探索