Pixels alone can carry reasoning: models trained from scratch on raw pixels learn to work problems step by step like language models
Synopsis
Existing image, video and unified multimodal generators almost never answer questions written in pixels on a canvas; they succeed only when a prompt rewriter or text module has already supplied the answer. To test whether pixels or training is at fault, researchers trained a family of purely pixel-space flow-matching generative models from scratch with no text pathway, from 7.6M to 458M parameters, taking and producing only pixels, reading questions, writing intermediate steps and answering on the canvas. These models show three hallmarks of large language models: per-step error falls with scale along a power law, frame-by-frame intermediate steps clearly beat single-frame answering, and reinforcement-learning post-training further strengthens reasoning.
Interpretation
The generation modules of image editing, video generation and unified multimodal models cannot perform symbolic reasoning over text on the canvas; they write correct answers almost only when a prompt rewriter or text module has already supplied the answer in text tokens. Prior generators render long legible passages and follow editing instructions, appearing close to reasoning over pixels; in the default condition the open image editing models FLUX.3-klein, Qwen-Image-2.1 and LongCat-Image-Edit score 0% on every task, including Copy where the answer is already printed in the question. With a rewriter, Qwen-Image-2.1 rises to 86.0% average accuracy and MiniMax-H3 to 98.3%; but deleting only the answer from the rewritten prompt while keeping everything else drops every tested generator to near 0%.
With the text pathway removed, pixel-only models read questions, write intermediate steps and produce answers directly on the canvas, reproducing the scaling law, chain of thought and reinforcement-learning gains of language models. No purely pixel-space generative model without text tokens had been systematically studied for these phenomena; direct answering stays near 1% on Word Problems at every size. On text tasks per-step error falls from 0.174 at 7.6M parameters to 0.037 at 458M, and a power law fitted only up to 130M predicts the held-out 458M error within 2%; at 130M, frame-by-frame reasoning solves 98.1% of Compute Then Paint and 49.2% of Trace Then Sum, against 31.1% and 17.2% for direct answering trained four times longer.
The model is a flow-matching transformer operating directly on raw pixels, with no VAE, tokenizer or text encoder, predicting the clean image and trained with the velocity loss following JiT. Frame-by-frame reasoning maps the canvas at step t to the canvas at step t+1; at test time frames are generated sequentially from the original problem canvas, each fed back in raw pixel space as the next input, so the canvas is the sole memory. Training covers four text tasks (Expressions, Multiplication, Word Problems, Factual Questions) and six cross-modal tasks (Mazes, Sum Mazes, Curve Tracing, Trace Then Sum, Count Then Paint, Compute Then Paint), evaluated on held-out problems never seen in training.
Reinforcement-learning post-training lets the model discover better reasoning trajectories from its own rollouts rather than imitating demonstrations. Supervised fine-tuning with the same compute yields limited gains, and non-optimal SFT barely helps. At 130M parameters GRPO raises Word Problems accuracy from 23.7% to 84.6%, while supervised fine-tuning with the same compute reaches at most 39.0%; RL surpasses optimal-demonstration SFT on all six tasks, for example 43.7% versus 35.7% on Trace Then Sum and 73.2% versus 61.0% on Compute Then Paint.
Perspective
The conclusions rest on ten tasks, four text tasks (Expressions, Multiplication, Word Problems, Factual Questions) and six cross-modal tasks (Mazes, Sum Mazes, Curve Tracing, Trace Then Sum, Count Then Paint, Compute Then Paint), and on five model sizes from 7.6M to 458M parameters, evaluated on held-out problems unseen in training; the diagnosis of existing generators covers image editing, video generation and unified multimodal models whose pipelines can be inspected.
The held-out scaling check is done only at 458M: measured text-task error is 2% below prediction, while the joint Curve Tracing model measures 8% above it; the frame-by-frame advantage appears only once the model is large enough to execute each step reliably, and Word Problems frame-by-frame accuracy is just 3.7% at 7.6M.
