Ouroboros (Self-Correcting Sampler) 🐍

Ouroboros is the self-correcting sampler: a closed loop where an LLM expands your prompt, an image gets sampled, a vision critic scores it and tells you (in the form of advice) how to improve the prompt, and it repeats until the score hits a target or you run out of iterations. It's one node that replaces the "generate, squint, retype the prompt, generate again" ritual you've been doing by hand - and it keeps the best image across the whole run, not just the last one.

How the loop works

Each iteration is: LLM rewrites/expands the prompt → an image is sampled → a vision critic scores it and returns concrete advice plus negative-prompt terms for the flaws it sees → the advice feeds the next revision. Two details make the loop meaningful rather than random:

• Fixed seed, fixed start latent. Only the prompt varies, so the score differences are a clean signal about the writing, not about seed luck. (Sampler Settings can opt out of this with its seed_mode, but fixed is the honest default.)
• Negative terms accumulate. The critic's flaw terms get deduped and appended onto your base negative each iteration, so the model is progressively told what not to do.

The loop is a plain Python while inside the node - no graph expansion - and generation runs through ComfyUI's own custom-sampler API (comfy.sample.sample_custom), which is what lets Sampler Settings expose per-sampler knobs like eta. Single-stage results match the classic KSampler path, so it's not some exotic sampler; it's a KSampler with a critic standing over its shoulder.

Wiring it up

The required inputs read like a shopping list of the pack's bundle idiom:

• model / clip / vae - your checkpoint.
• latent - an Empty Latent matching your model sets the image dimensions. For img2img or hires feedback, feed a real latent and set denoise < 1 in Sampler Settings.
• user_prompt - your intent. The enhancer improves it; the critic anchors its advice to it. negative is the base; flaw terms accumulate onto it.
• enhancer_settings - a Local LLM Settings (GGUF) for the prompt LLM. Its system prompt defines how the prompt is expanded. On low VRAM, point it at the same model file as the critic.
• critic_settings - an Ouroboros Critic Settings bundle: the vision judge + scoring rules.
• sampler_settings - a Sampler Settings bundle, or a chain of them. One = classic single pass. Two or more = refine mode: stage 1 drafts, later stages polish the previous latent (give them denoise < 1), and the critic judges only the final image.
• target_score and max_iterations - the two stop conditions. The score scale comes from the critic's score_min..score_max (default target 4 out of 5-ish on that node's default rubric), so set it on the Critic Settings, not here.
• low_vram - on by default, and leave it on unless you have VRAM to spare. It keeps strictly one model in VRAM at a time: frees ComfyUI models before each LLM call, unloads the worker after, and auto-sizes each LLM's context to what the request needs.

Useful optional: trigger_words (comma-separated, e.g. LoRA triggers from Lora Unlim Accumulator) are always appended to the enhanced prompt so an LLM rewrite can't drop them. enhancer_history recaps the last N prompt+score pairs as "already tried, don't repeat." model_negative is for checkpoints shipped as a model + uncond-model pair (Ideogram) - note it does nothing at cfg 1.0, because the guider skips the uncond pass there.

Outputs

images, prompts, judge_data, captions, times and settings_data all feed Image Compare directly - wire the whole trajectory in and flip through iterations. best_image / best_prompt / best_score are the winners. report has the full trace. Wire best_score into a Settings Save and your measured results start building a library of what actually works.

Installing it

Same pack: ComfyUI Manager → "Kinburg-Nodes", or git clone https://github.com/Kinburg/Kinburg-Nodes into ComfyUI/custom_nodes, restart. You need two model downloads of your own: a GGUF LLM for the enhancer and a vision (mmproj) model for the critic, plus llama-cpp-python - installed automatically by the pack's install.py. First-time confusion is almost always the critic: a vision model without an mmproj isn't a critic, it's a chat model looking at nothing.