Generative models increasingly write a chain of thought before drawing, and report large gains from doing so. We ask what that text contributes, by substitution rather than ablation: holding the generator, sampler and seeds fixed, we replace the think block with controlled alternatives and measure the image. Across a unified model with a think mode, a reinforcement-learned reasoning generator, and a planner driving a frozen diffusion model, the same picture emerges. Content-free prose of matched length is worth nothing or less (−0.05 to −0.15). One sentence from an unrelated 8B model, naming the subject and a few identifying details, matches the generator's own plan in three to six times fewer conditioning tokens. The value sits where the referent is in doubt: writing down a name is worth +0.33 when it is right and +0.05 when it is wrong, and where the generator already succeeds the whole effect is −0.07±0.07. Cutting a plan at the text encoder's window leaves 272 of 286 images byte-identical, and a plan rewritten to fit that window ties a bare identification (+0.003±0.044), so a plan contains nothing the identification does not already supply. Field-level interventions find only colour executed reliably, and a control writing the same sentences into the user prompt shows that limit belongs to the generator rather than to the planning channel. Two judges and two annotators replicate the central contrasts. A think block earns its keep as a specification of what to draw, not as a plan that is carried out.
Figures & tables
think block
words
n
score
gain
none
0
215
0.87
—
length-matched filler
139
215
0.82
−0.05±0.04
model’s own reasoning
131
215
1.27
+0.40±0.06
model’s own name
8
175
1.07
+0.20±0.05
external name (8B LLM)
8
200
1.07
+0.18±0.04
external description (8B LLM)
30
200
1.24
+0.34±0.05
Table 1: WISE consistency by arm (0–2 scale). Gains are paired differences against the no-plan arm. Seeds of the same prompt are not independent, so each prompt’s difference is averaged over its seeds first and n counts prompts, not images; the standard errors are therefore clustered at the prompt level.
Figure 1: (a) WISE consistency for each substitute think block, with standard errors; the dotted line is the no-plan baseline. (b) The same scores against the length of the think block: quality is set by what the block says, not by how long it is.
Figure 2: Same generator, same prompt, same seed, four think blocks. Filler of the same length as the model’s own reasoning leaves the image where the no-plan arm left it. Reasoning helps when the model names the referent correctly (row 1, raccoon) and not otherwise: it writes its way from “a bird embodying French heritage” to a spaniel in a palace (row 2), and it keeps the daylight that the time zone forbids at the Brooklyn Bridge (row 3). The reference description fixes both.
stratum
none
own name
own reas.
reference
right ( n=92 )
0.94
1.27 ( +.33 )
1.49 ( +.55 )
1.61 ( +.67 )
wrong ( n=83 )
0.80
0.86 ( +.05 )
1.01 ( +.21 )
1.36 ( +.56 )
Table 2: WISE consistency within strata defined by the model’s own identification, a variable fixed before any think block is chosen. Parentheses give the paired gain over the no-plan arm in the same stratum; standard errors of these gains, clustered at the prompt level, are 0.05 to 0.10 .
unaided
n
none
filler
own
reference
fails (0)
89
0.27
0.31 ( +.04 )
0.88 ( +.61 )
1.26 ( +.99 )
partial (1)
51
0.98
0.71 ( −.27 )
1.39 ( +.41 )
1.49 ( +.51 )
succeeds (2)
60
1.88
1.73 ( −.15 )
1.82 ( −.07 )
1.88 ( +.00 )
Table 3: WISE consistency within strata. Each prompt is assigned to a stratum by its no-plan score on seed 1 and every arm is read on seed 2, so each prompt appears once. Parentheses give the paired gain with its standard error.
BAGEL-7B-MoT
T2I-R1
Qwen3-8B + SDXL
think block
gain
words
gain
words
gain
words
no plan
—
0
—
0
—
0
length-matched filler
−0.05±0.04
139
−0.10±0.04
111
−0.15±0.04
109
the model’s own plan
+0.40±0.06
131
+0.27±0.05
105
+0.18±0.04
112
external name (8B)
+0.18±0.04
8
+0.16±0.04
19
+0.14±0.04
19
external sentence (8B)
+0.34±0.05
30
+0.41±0.05
41
+0.30±0.05
41
Table 4: The same decomposition on three generators. Each entry is a paired difference against that generator’s own no-plan arm, with the mean length of the substituted block. The ordering of the arms is identical in all three.
field
B plan
B edit
B par.
R plan
R edit
R par.
S plan
S edit
S par.
colour
+.58
+.67
−.10
+.53
+.75
−.37
+.45
+.53
−.17
count
+.17
+.10
−.02
+.15
+.12
−.25
+.10
−.02
−.10
spatial
+.02
+.20
−.18
+.07
+.43
−.10
−.07
−.02
−.02
absence
−.13
+.37
−.15
±.00
+.57
−.08
−.12
+.27
−.13
Table 5: Field-level transmission on both architectures, 20 scenes × 3 seeds ( n=60 per cell). Plan is compliance when the plan states the field minus compliance when it is silent; edit is the probability the image takes the new value after one field is changed, minus the same probability under the unedited plan; para. is the change in compliance under a meaning-preserving paraphrase. Columns prefixed B are BAGEL-7B-MoT, R is T2I-R1 and S is SDXL.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
contrast
humans
judge
diff.
filler − no plan
+0.08±0.10
−0.10±0.09
1.3σ
own reasoning − no plan
+0.56±0.14
+0.33±0.15
1.2σ
reference − no plan
+0.59±0.12
+0.70±0.15
0.6σ
reference − own reasoning
+0.03±0.11
+0.38±0.11
2.2σ
Appendix
Table 6: Round two. Paired gains over the no-plan arm on the same 40 prompts, by two human annotators and by the judge. The first three rows replicate; the fourth does not.
system
arm
judge 1
judge 2
T2I-R1
filler
−0.10±0.04
−0.06±0.05
its own plan
+0.27±0.05
+0.25±0.05
external sentence
+0.41±0.05
+0.25±0.05
reference
+0.76±0.05
+0.54±0.06
sentence − own plan
+0.13±0.06
+0.00±0.05
planner
filler
−0.15±0.04
−0.17±0.04
Appendix
Table 7: Every arm of the two further pipelines, scored by both judges. The ordering is preserved throughout, but the advantage of the external sentence over the generator’s own plan is not.
Figure 3: Minimal edits of a single plan field, same generator and seed. Changing the colour word changes the image; changing the count from two to five does not. In both reference plans the one object the plan forbids is present anyway: a lit candle on the table that is to have no candles, a toothbrush on the shelf that is to have none.
field
D silent
A in the plan
P in the prompt
colour
0.22
0.80 ( +.58±.08 , p<.001 )
0.95 ( +.73±.09 , p<.001 )
count
0.18
0.35 ( +.17±.08 , p=.06 )
0.43 ( +.25±.11 , p=.04 )
spatial
0.55
0.57 ( +.02±.10 , p=.87 )
0.70 ( +.15±.11 , p=.19 )
absence
0.83
0.70 ( −.13±.07 , p=.09 )
0.73 ( −.10±.11 , p=.38 )
Appendix
Table 8: The same four sentences in the plan (A) or in the user prompt (P), against the silent control (D). Compliance on the unified model, 20 scenes × 3 seeds; differences are paired on scene, with scene-clustered standard errors and two-sided p values over the 20 scenes.
Figure 4: Field-level transmission, 20 scenes × 3 seeds ( n=60 per cell). (a) Compliance when the plan states the field minus compliance when it says nothing about it. (b) Probability that the image shows the new value after exactly one field of the plan is edited, minus the same probability under the unedited plan.
category
n
none
filler
own reas.
own name
LLM name
LLM descr.
ref.+elab.
reference
Biology
50
0.72
0.68 ( −.04 )
1.14 ( +.42 )
1.02 ( +.24 )
0.98 ( +.25 )
1.08 ( +.35 )
1.31 ( +.47 )
1.48 ( +.76 )
Chemistry
51
0.49
0.39 ( −.10 )
0.80 ( +.31 )
0.48 ( −.02 )
0.50 ( −.02 )
0.77 ( +.25 )
1.14 ( +.61 )
1.27 ( +.78 )
Cultural knowledge
166
0.95
0.93 ( −.02 )
1.51 ( +.56 )
1.26 ( +.32 )
1.28 ( +.31 )
1.46 ( +.49 )
1.43 ( +.50 )
1.60 ( +.64 )
Physical knowledge
49
0.98
0.73 ( −.24 )
1.14 ( +.16 )
1.02 ( +.10 )
1.06 ( +.10 )
1.10 ( +.15 )
1.53 ( +.53 )
1.53 ( +.55 )
Space
50
1.10
1.14 ( +.04 )
1.36 ( +.26 )
1.21 ( +.12 )
1.10 ( −.02 )
1.33 ( +.21 )
1.64 ( +.47 )
1.48 ( +.38 )
Time
49
0.90
0.88 ( −.02 )
1.14 ( +.24 )
1.00 ( +.14 )
1.02 ( +.12 )
1.15 ( +.25 )
1.06 ( +.28 )
1.53 ( +.63 )
Appendix
Table 9: WISE consistency by category and arm. Each cell is the arm’s mean on that category with, in parentheses, its paired gain over the no-plan arm on the cells the two arms share. n counts no-plan images in the category.
task
none
own reas.
filler
restate
single object
0.950
0.964
0.971
0.971
two objects
0.904
0.924
0.866
0.927
counting
0.447
0.474
0.461
0.474
colours
0.931
0.941
0.915
0.952
position
0.848
0.914
0.798
0.874
all
0.880
0.904
0.865
0.903
Appendix
Table 10: GenEval accuracy by task. All arms cover an identical set of 510 cells over 335 prompts.
field
B early
B late
B silent
S early
S late
S silent
colour
0.80
0.95
0.22
0.67
0.20
0.22
count
0.35
0.50
0.18
0.28
0.18
0.18
spatial
0.57
0.50
0.55
0.45
0.48
0.52
absence
0.70
0.65
0.83
0.70
0.82
0.82
Appendix
Table 11: The late-position arm. The same four sentences are moved behind 71 words of generic material, with their content unchanged. For the frozen generator every field returns to the value it takes when the plan is silent; for the long-context unified model none of them does. Columns prefixed B are BAGEL-7B-MoT and S are SDXL.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Reasoning-augmented text-to-image models such as GoT-R1 emit an explicit textual plan - object names, attributes, and bounding boxes - before generating image tokens. When such a model fails a compositional prompt, is the plan wrong, or is the plan right and the decoder unfaithful? Because the plan is machine-readable it can be edited before decoding, which makes the two separable. We first validate the ruler. Swapping the two bounding boxes inside the model's own chain demonstrably flips the generated layout: detector-based accuracy falls 0.75 -> 0.48 (p<1e-3), while a widely used VQA-based spatial metric rises. A five-rater human study agrees with the detector on 81% of items and with the VQA judge on 57%. All spatial results therefore use geometric scoring. Under sound measurement the decoder is a faithful executor: 94% of generated layouts realize the planned relation, and object-box binding survives reordering of the plan's object segments. The planner is the bottleneck. It writes wrong relations for phrasing-dependent reasons - 98% accuracy on "left" against 54% on "right" for semantically identical layouts, a raster-order bias we isolate with a mention-order control - and cluttered geometry that the decoder faithfully reproduces. Editing the plan therefore fixes the image without retraining: symbolic verification with resampling gives +5.0 points (p<1e-3), minimal in-place repair +6.0 (p=.02), rewriting only box geometry +10.7 (p<1e-4), and replacing the plan outright +13.3 (p=1e-4). Gains are indifferent to the plan's prose style and to its likelihood under the planner, but not to its geometry. Modular planner-decoder designs are therefore viable, provided the plan is internally consistent: box-text contradictions induce object duplication and identity fusion. We release the plan-fidelity evaluation protocol, all plans, and 12k generated images.
Diffusion large language models (dLLMs) generate text via iterative denoising but consistently underperform on multi-step reasoning. We hypothesize this gap stems from a coordination problem: AR models build coherence token-by-token, while diffusion models must coordinate all positions simultaneously. We propose plan conditioning, a training-free method that prepends a short (~100-token) natural-language plan from an AR model to the diffusion model's prompt. The plan serves as a frozen scaffold -- globally visible context that every token position can attend to from the first denoising step. On GSM8K, plan conditioning improves LLaDA-8B-Instruct from 75.6% to 87.2% (+11.6 percentage points), matching a same-size AR model (LLaMA 3.1 8B, 87.7%) despite a 6.4pp weaker baseline. On HumanEval, the gain is +12.8pp (37.2% to 50.0%), showing plans generalize to code. The same plans improve LLaMA by only +5.7pp on GSM8K and +1.3pp on HumanEval -- diffusion models benefit 2-10x more, supporting the coordination-problem hypothesis. Across 5 random seeds, plan-conditioned GSM8K accuracy has zero standard deviation, making diffusion inference highly stable. Ablations reveal the model follows plan strategy (wrong-strategy plans cause -16.3pp) but is robust to plan values (perturbed numbers: -1.1pp), and that planner quality has a sharp threshold: smaller Llama-class plans hurt (-1.6 to -6.8pp) while frontier plans provide the full lift. Attention analysis confirms the mechanism: plan tokens receive 1.8x excess attention during early denoising, declining to uniform as completion tokens solidify. Plan conditioning costs ~$0.002 per problem and adds ~2s of latency.
Diffusion models generate realistic images but often fail on visual reasoning tasks, such as filling in a Sudoku or drawing the path through a maze. When a discrete symbolic representation is available, recursive methods such as the Tiny Recursive Model (TRM) solve even hard instances of these puzzles. We ask how such reasoning can be carried over to pixels, where no symbolic representation is available. We propose Painter-Thinker (PaTh): a small recursive network (the Thinker) reasons over a grid of learned tokens that encode the noisy image and the conditioning, refines a latent state within every denoising step, and steers a frozen diffusion model (the Painter) through ControlNet adapters. The Thinker is trained with the standard reconstruction loss alone, without symbolic targets, a solver, or a verifier. PaTh solves 92.5% of hard MNIST Sudoku puzzles (prior best 75%) and 71.2% of extreme ones (prior best 4.1%), with 10M parameters against 82M for a standard diffusion model. It also improves on mazes, Queens, and CLEVR scenes with specified spatial relations, and its advantage grows with problem size. Diagnostic experiments show that PaTh recovers from injected mistakes that the diffusion model cannot repair, especially when many cells are wrong. Together, these results show that reasoning mechanisms developed for symbolic data can be integrated into pixel-space diffusion without symbolic supervision, opening a path toward generating data under increasingly complex constraints.
Paweł Skierś, Małgorzata Grzanka, Wojciech Masarczyk +2
Warsaw University of Technology · IDEAS Research Institute · University of Amsterdam