A fixed image pipeline.
An agent making the decisions.
Three experiments in iterative image generation with ComfyUI. The workflow never changes; between frames, an LLM agent scores the image against a fixed rubric and writes the next prompt. The same grey coupe runs through all three: refined into a studio hero shot, moved from the studio onto a coastal highway through chained img2img, and improved further by selection and mutation across a population of prompts and seeds.
How the work is divided
Everything that repeats is code
A small script watches a folder. When a prompt file appears, it runs it through the same frozen ComfyUI graph, saves the image, and appends one line to a log. It makes no judgment calls and never changes between runs, so results are repeatable and failures are easy to isolate.
Everything that requires judgment is the agent
Between frames, an LLM looks at the image, scores it against a rubric locked before the run started, and writes the next prompt as a small JSON file. The two halves communicate only through files on disk, so the agent can be swapped — a frontier model, a local model, or a person with a text editor — without changing anything else.
The three experiments
Hill-Climb
Morph
Evolve
What the logs taught us
- Seed handling depends on the loop type. Fix the seed when refining a text-to-image prompt, so each change is attributable to the edit. Rotate it in img2img chains, where a fixed seed re-applies the same noise every frame and locks artifacts in. In evolution, treat the seed as one more variable to search.
- Low denoise preserves artifacts rather than repairing them. When a chain degrades, roll back to a clean frame and regenerate at higher denoise. Experiment 02's collapsed lineage is preserved in full.
- Persistent compositions respond to relabeling, not force. A garage doorway that held through every denoise setting changed immediately once the prompt described the same shape as a highway underpass.
- Describe the cause, not the desired result. "Fine metal flake catching the light" produced the highest-scoring paint of the project; specifying the exact reflection shapes degraded the image every time it was tried.
The full write-up is on the findings page.
Every prompt, seed, score and judge note is in data/, unedited.