How Far Can Coding Agents Go in Visual Generation?

How Far Can Coding Agents Go in Visual Generation?. By Hengguang Zhou and five others, from the University of California, Los Angeles and Tsinghua University. Published on the thirteenth of October. No code repository is linked to it.

Modern image generation mostly relies on specialized generative models, and those models leave the process of building an image implicit, buried in learned parameters. This study explores an alternative: general-purpose coding agents that construct images by writing executable programs. The claim is that such agents show a distinctive capability profile, strong on structure and precise constraints, weaker on perceptual richness. And beyond that, the authors argue the agents seem to hold a consistent, coherent picture of a scene before any drawing program comes into play.

The first piece of evidence is zero-shot drawing on AgentDrawBench, a benchmark of two hundred prompts. Agents excel at preserving structure and at satisfying precise visual constraints. They reach competitive results in domains where compositional accuracy matters more than photorealistic appearance, while remaining limited in perceptual richness. How expressive they are visually depends on the drawing interface they are given. A qualitative study on reconstruction adds that procedural skills, extracted from human demonstrations, support different artistic media.

The second piece treats code as an external visual language of thought: a structured, executable representation that an agent can both construct and inspect. This is tested on visual question answering. Visual code reconstructed from an image retains substantial information for answering questions, and in some domains it supports higher accuracy than the original image does.

The third piece goes further inward. If agents can do all this, do they draw on a shared internal representation of scene content, one that sits beyond any particular drawing program? The authors explore what they call an internal visual language-of-thought hypothesis: a shared, compositionally structured representation that supports visual operations across drawing languages. Two observations bear on it. Drawings made through different harnesses, and images assembled from independent single-pixel queries, show related layouts and tones. And agents keep self-generated room arrangements consistent across viewpoints even after the visible drawings have been removed.

The scope deserves care. The drawing results come from one benchmark of two hundred prompts, in a zero-shot setting. The reconstruction study is qualitative. The advantage over the original image in question answering appears in some domains, not all. And the internal representation is framed as a hypothesis the authors explore, supported by layout and tone similarity and by consistency across viewpoints, rather than something observed directly. And where photorealistic appearance is what counts, perceptual richness remains the agents' limit.

Read together, the three strands move from output to representation. Agents drawing in code are good at structure, the code they write can stand in for an image when answering questions, and the agreement across harnesses and viewpoints is taken as a sign that a consistent world is already being held in mind before the harness gets involved and generation begins.