Text, images, audio, touch, and even robot actions can, in principle, all be rendered onto a single canvas, suggesting an ambitious goal of unifying all modalities within vision. Language provides the natural first test. To subsume it, a visual model must reason directly over text in pixels. Across a wide range of image, video, and unified multimodal generators, however, we find that their visual generation modules cannot perform symbolic reasoning over text on the canvas. They succeed only when an underlying text module or prompt rewriter already supplies the answer, effectively reducing the generator to a mere renderer, a limitation that post-training does little to resolve. To test whether this failure stems from pixels themselves or from current training paradigms, we train a family of purely pixel-space generative models from scratch without any text pathway. With images as their sole input and output, these models read questions, write intermediate steps, and produce answers directly on the canvas. Remarkably, they exhibit hallmark behaviors of large language models. Performance improves predictably with scale following a power law. Writing intermediate steps frame by frame substantially outperforms single-frame answering, echoing chain-of-thought, while reinforcement-learning post-training further strengthens reasoning. Moreover, sharing one canvas allows the same model to reason across modalities, a capability that text-only language models lack. To the best of our knowledge, this is the first systematic study of these phenomena in purely pixel-space generative models free of text tokens.
Figures & tables
Figure 1: Unifying modalities and reasoning in pixel space. (a) In our envisioned unified paradigm, text, images, audio, touch, and actions share one canvas, where a pixel-only model reads the current state, generates the next canvas, and feeds it back as input. (b) Current generators rely on text tokens. Such a generator typically writes the answer only when explicitly supplied with it, as shown by real outputs from Qwen-Image-2.1. (c) Trained from scratch without text tokens, our pixel model reasons across modalities frame by frame. Real rollouts from a 33M model on unseen problems interleave arithmetic, dot counting, and region painting, or curve tracing and summation.
Figure 2: The six tasks of the benchmark. Ordered by decreasing answer visibility from string copying to factual recall, each task tests whether a generator can read a question on the canvas and write the answer into the empty box.
Model
Default
Rewrite
Delete
Render
Think
Image editing models
GPT-Image-2.5
–
98.0
–
–
–
Gemini-3-Pro-Image
–
94.1
–
–
–
Nano-Banana-2.1
–
92.0
–
–
–
Qwen-Image-2.0-Pro
0.0
91.7
–
–
–
Qwen-Image-2.1
0.0
86.0
0.0
–
–
Table 1: Exact accuracy in percent of current visual generators. In Default , models receive only the canvas and a fixed prompt without the answer. For image and video models, Rewrite uses a prompt rewriter that often supplies the answer, and Delete removes the answer from the rewritten prompt. For unified models, Render writes the answer into the prompt as text tokens, and Think turns on thinking mode, enabling text-token reasoning before generation. Video models are judged on the final frame. Dashes mark inapplicable or unavailable conditions, as closed systems whose prompt rewriter cannot be disabled lack Default, and systems whose internal rewritten prompt cannot be edited lack Delete. The highest value per row is in bold.
Figure 3: Test-set accuracy after post-training the generation module of four image editing models. All base models start at 0% on every task. Post-training teaches the generators to copy visible strings, yet accuracy falls steeply as answer visibility in the question decreases and stays at 0% to 2% on Arithmetic and Factual.
Figure 4: Overview of the ten pixel-space reasoning tasks. Each 192 × 192 (256 × 256 for Word Problems) canvas places the question and correct intermediate reasoning steps above a horizontal rule, with the final answer written below.
Figure 5: Pixel-space reasoning obeys power-law scaling. From 7.6M to 458M parameters, per-step error in generating the next frame falls linearly on log-log scales for two models, one evaluated across Expressions, Multiplication, and Factual Questions and the other on Curve Tracing. Power laws fitted to medians from 7.6M to 130M extrapolate accurately to the held-out 458M model, with dashed lines matching measured medians within 2% and 8%. Across five seeds per size, boxes cover the middle three, whiskers span all five, and black bars mark medians.
Figure 6: Frame-by-frame chain of thought and reinforcement learning at 33M and 130M. (a) Writing steps frame by frame outperforms direct answering across all six tasks at equal training of 40k steps and surpasses four times longer direct training on four of them. (b) On the four cross-modal tasks, each run continues from a Start model trained on non-optimal demonstrations with extra steps or detours. Under equal compute, RL outperforms non-optimal SFT and surpasses optimal SFT on four tasks at 33M and all six at 130M. We apply GRPO to the two text tasks and RFT to the four cross-modal tasks. For the four multimodal tasks, test accuracy requires a correct answer in the fewest frames.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
System, condition
Factual (paraphrase)
Factual (unseen)
Arithmetic
Instruction
Context
Visual
Copy
Image editing models, Rewrite (Default is 0 on every set for all six models)
FLUX.2-klein, official prompt upsampler
57
57
73
53
45
60
43
FLUX.3-klein, official prompt upsampler
92
90
78
63
77
87
97
Qwen-Image-2.1, API rewrite
95
90
92
92
95
95
43
LongCat-Image-Edit, upsampler-style rewrite
37
45
25
5
17
12
2
Qwen-Image-2.0-Pro, its own rewriting
100
–
90
100
100
100
60
Appendix
Table 2: Table 1 per test set, as exact accuracy in percent. Instruction holds new words under trained rules. With the answer deleted from the rewrite, all four open image editing models score 0 on Factual and Arithmetic. In SenseNova’s Think mode, the image is correct on 99.0% of the items whose thought contains the answer.
Video generation model
Default, final frame
Default, any of 8 frames
Rewrite
Delete
Wan2.2
0
0
0.7
0
OmniWeaving
0
0.2
24.0 (any frame: 28.8)
0
Cosmos3
0
0
13.6 (any frame: 17.4)
0
MiniMax-H3
1.4
1.4
98.3
0
Wan 3.0
0
1.7
83.3
–
Appendix
Table 3: Video generation models, as exact accuracy in percent. Default gives the canvas as the first frame and a fixed prompt without the answer, judged on 8 evenly sampled frames. In 88% to 98% of the frames of Wan2.2, OmniWeaving, and Cosmos3 the answer box stays empty. Delete is evaluated on Factual and Arithmetic, where the answer cannot be read off the canvas. For Wan 3.0 the rewrite is its own, with prompt_extend on as described in Appendix B .
System (official interface)
Rewriter
All tasks
Copy
Instruction
Context
Visual
Arithmetic
Factual
Qwen-Image-2.0-Pro (image editing)
off
0
0
0
0
0
0
0
on
91.7
60
100
100
100
90
100
Wan2.6-Image (image editing)
off
0
0
0
0
0
0
0
on
81.7
30
90
80
100
90
100
Wan 3.0 (video generation, final frame)
off
0 (any frame: 1.7)
0
0
0
0
0
0
on
83.3 (any frame: 91.7)
90
80
90
80
70
90
Appendix
Table 4: One switch with everything else fixed. The same canvas, the same fixed prompt, the same seed, and the same model are used, with the model’s own prompt rewriting, prompt_extend, turned off or on. Accuracy is exact and in percent. Wan 3.0 starts from the canvas as its first frame, generates 5 s of video, and is judged on 8 sampled frames. With the rewriter off, the two image editing models leave the answer box empty on every item, and Wan 3.0 leaves it empty or writes scribbles such as “17” or “#ab)”.
Task, set
FLUX.2-klein
FLUX.3-klein
Qwen-Image-2.1
LongCat-Image-Edit
Copy, new strings
79.5
92.5
91.5
81.0
Instruction, new words / new rule
10.0 / 0
81.7 / 0
11.7 / 0
10.0 / 0
Instruction, training items
0
83.0
0
0
Context, new names
8.0
28.5
9.0
0.5
Context, training items
8.0
32.0
5.0
0
Visual, new scenes
6.5
7.0
3.5
8.0
Appendix
Table 5: The numbers of Figure 3 , for post-training of the generation module, as exact accuracy in percent on the test sets with an MLLM judge. Every base model scores 0 on every row. The Instruction row gives new words and a rule never trained. The training-item rows test recall on 100 trained items per task, each seen about 19 times, and match the test rows, so post-training teaches copying of visible text rather than reasoning.
Size
XS
S
M
B
L
Width / depth / heads
256 / 6 / 4
384 / 12 / 6
512 / 12 / 8
768 / 12 / 12
1024 / 24 / 16
Parameters
7.6M
33M
58M
130M
458M
Appendix
Table 6: Model sizes. All sizes use the same recipe, with batch 256, 128 on Word Problems , AdamW with β of 0.9 and 0.95, learning rate 10 − 4 with 1,000 warm-up steps, EMA 0.999, and bf16, for 40k steps in Sections 6 and 7 and 132k steps in the scaling ladders of Section 5 .
Curve
Fit on 7.6M to 130M
Log RMSE
458M predicted (90% interval)
458M measured (90% interval)
Relative error (90% interval)
Fit on all five sizes
Text tasks
49.6 ⋅ N − 0.360
0.056
0.0381 (0.0337–0.0443)
0.0374 (0.0271–0.0442)
− 2% ( − 26% to +23%)
52.6 ⋅ N − 0.363 (0.050)
Curve Tracing , joint model
111 ⋅ N − 0.340
0.052
0.1259 (0.1126–0.1335)
0.1359 (0.1234–0.1559)
+8% ( − 4% to +24%)
84.0 ⋅ N − 0.324 (0.051)
Appendix
Table 7: Power-law fits of the per-step error of Section 5 , on the medians of five seeds per size. The last column fits all five sizes and gives the log RMSE after the fit.
Large language models (LLMs) can generate visual assets through text-based formats such as SVG and CAD, but these symbolic representations describe images using high-level primitives rather than explicit raster values. We ask whether LLMs can instead generate pixel images directly through their existing vocabulary, without a learned image tokenizer or decoder. We introduce PixelBench, a benchmark for text-to-pixel (T2P) and image-to-pixel (I2P) generation across multiple resolutions, together with a complementary pixel-understanding evaluation. Models generate an RGB color palette followed by a grid of indices specifying the palette color of each output pixel. In T2P, the strongest models produce recognizable images directly from text. In I2P, the strongest models outperform palette-matched downsampling and specialized pixelization algorithms at low resolutions. Pixel-understanding performance correlates strongly with generation performance. A resolution-scaling study shows that pixel-generation performance often improves initially before declining at high resolutions. Finally, we introduce Pixel Reconstruction Tuning (PRT), which augments visual instruction tuning with caption-palette-grid trajectories. We post-train Molmo2-4B from its pretraining checkpoint by mixing general instruction data with our curated PRT data, improving performance across 16 visual-centric benchmarks, indicating that explicit pixel reconstruction can provide useful auxiliary supervision.
Pixel art is widely used in games and digital content creation, requiring readable forms, coherent spatial relationships, and harmonious colors under limited resolution and palette constraints. Language models show growing promise in visual creation, yet their pixel art creation capabilities lack systematic evaluation. We introduce PixelBench, formalizing text-to-image generation and instruction-guided editing for pixel art and providing a benchmark for these tasks. We further present a creation environment in which language model agents invoke drawing tools through code to generate and edit pixel art. We assess the quality of generated artwork through human evaluation and employ a multimodal large language model as a judge, validating the agreement between its automated assessments and human preferences. We compare language models and configurations across diverse subjects and canvas sizes. In our experiments, GPT-6 Astra achieves both the highest average judge score and the best average human rank, highlighting the potential of general-purpose language models for pixel art creation. We release the tasks, all artworks, the human rankings, and the judge protocol.
Mutian Hong, Jianxuan Yao, Mingkun Yue +3
Westlake University · Mind Lab · ShanghaiTech University
Unified Multimodal Models (UMMs) often rely on separate visual representations for understanding and generation, increasing visual context length and complicating integration with established vision-language pretraining pipelines. Recent advances in pixel-space modeling offer an encoder-free alternative, but extending this paradigm from images to videos is non-trivial: video understanding and generation adopt different temporal representations, leaving the design of a unified visual interface an open question. We present PixelUMM, an encoder-free model for unified image and video understanding and generation directly in pixel space. PixelUMM represents images as spatial patches and videos as spatiotemporal tubelets, connecting raw pixels to a shared multimodal backbone through single-layer linear projections. Its Mixture-of-Transformers architecture combines shared attention with task-specific parameters and extends clean-pixel prediction to video generation, jointly supporting autoregressive text prediction and pixel-space flow matching. Experiments show that PixelUMM achieves competitive performance across image and video understanding and generation tasks. We further conduct empirical studies of key design choices, including decoder design and spatial-temporal patch size, providing insights for future pixel-space unified multimodal models.