Ouroboros (Self-Correcting Sampler) 🐍

Self-correcting text→image loop: an LLM expands the prompt, an image is sampled, a vision critic scores it and advises how to improve the prompt, and it repeats until the target score (or max iterations). Outputs the whole trajectory (wire images/prompts/judge_data straight into Image Compare) plus the best image/prompt.

By Kinburg·Created 2 months ago·Updated 3 days ago· 1

model
clip
vae
latent
enhancer_settings
critic_settings
sampler_settings
model_negative

images
prompts
judge_data
best_image
best_prompt
best_score
report
iterations
captions
times
settings_data

◄user_prompt►

◄negative►

◄target_score4.0►

◄max_iterations6►

◄low_vramtrue►

◄enhancer_history4►

◄trigger_words—►

◄full_console_logtrue►

◄log_modestreaming►

CategoryKinburg-Nodes/Bestiary/Ouroboros

Inputs (17)

Name	Type	Default	Description
model	MODEL		—
clip	CLIP		—
vae	VAE		—
latent	LATENT		Starting latent — sets the image dimensions (wire an Empty Latent, matching your model: SD/SDXL vs SD3/Flux). For img2img/hires-feedback, feed a real latent and set denoise<1 in Sampler Settings. Reused each iteration so only the prompt varies.
user_prompt	STRING		Your intent — what you want. The enhancer improves it; the critic anchors advice to it.
negative	STRING		Base negative prompt; the critic's flaw terms accumulate onto it each iteration.
enhancer_settings	KINBURG_LLM_CONFIG		A 'Local LLM Settings (GGUF)' for the prompt-enhancer LLM — its model/temperature AND its system_prompt define how the prompt is expanded. On low VRAM, point it at the SAME model file as the critic.
critic_settings	KINBURG_CRITIC		A 'Ouroboros Critic Settings 🐍' bundle — the vision judge + scoring rules.
sampler_settings	KINBURG_SAMPLER_CFG		A 'Sampler Settings' bundle — OR a chain of them (wire Sampler Settings nodes in series). One = classic single pass. Two or more = REFINE mode: each iteration runs the stages in order (stage 1 drafts, later stages polish the previous latent with their own denoise / sampler / eta); the critic judges only the FINAL image.
target_score	FLOAT	4.00–100	The loop stops once the overall score reaches this (on the critic's score_min..score_max scale).
max_iterations	INT	61–100	Hard cap on refinement rounds.
low_vram	BOOLEAN	true	Low-VRAM mode (default ON). ON: strictly one model in VRAM at a time — frees ComfyUI models before each LLM call, unloads the LLM worker after it, and AUTO-SIZES each LLM's context (n_ctx) to what the request actually needs, so the n_ctx set on the LLM Settings nodes becomes the MAXIMUM. Best when the GPU can't hold the diffusion model + the LLM at once. OFF: keep everything resident (fastest, no reloads) and use the full n_ctx from the Settings nodes — for GPUs with VRAM to spare. NOTE: inside the loop this overrides the unload_* toggles on the LLM Settings nodes.
model_negativeopt	MODEL		Second model for the UNCONDITIONAL (negative) pass — for checkpoints shipped as a model + uncond-model pair, such as Ideogram. The loop builds the dual-model guider itself on every iteration, so the freshly rewritten prompt is still what gets encoded (wiring a ready-made 'Dual Model CFG Guider' node here instead would freeze the conditioning and defeat the whole loop). NOTE: at cfg 1.0 the guider skips the uncond pass entirely and this model does nothing — give the Sampler Settings a cfg above 1.
enhancer_historyopt	INT	40–50	How many of the most recent iterations (prompt + score) are recapped to the enhancer LLM as 'already tried, do not repeat' context. Higher = more memory of what failed (longer prompt, more tokens); 0 = no history (each rewrite sees only the intent, current prompt and the latest critic advice). The latest advice is always sent regardless.
trigger_wordsopt	STRING		Comma-separated words ALWAYS appended to the enhanced prompt (e.g. LoRA triggers from Lora Unlim Accumulator's 'triggers' output), so the LLM rewrite can't drop them.
full_console_logopt	BOOLEAN	true	Console/terminal verbosity only (not the Live Log node or 'report'). On: per iteration also prints the enhanced prompt, advice and negative additions. Off: just a one-line score per iteration. The full trace is always in the 'report' output and the Live Log node.
log_modeopt	COMBO	streaming	How the Ouroboros Live Log node updates: 'streaming' is like 'per step' but the enhancer's prompt types out token by token as it's written (the critic's verdict still arrives whole — it simply isn't wired to the log that way); 'per step' posts each stage the moment it finishes (enhanced prompt → generated image → critic verdict), each timestamped; 'per iteration' posts one combined entry after the whole iteration (the older behaviour).

Outputs (11)

Name	Type	Description
images	IMAGE	—
prompts	STRING	—
judge_data	STRING	—
best_image	IMAGE	—
best_prompt	STRING	—
best_score	FLOAT	—
report	STRING	—
iterations	INT	—
captions	STRING	—
times	STRING	—
settings_data	GEN_SETTINGS	—