Nodes/Kinburg-Nodes/Vision LLM Judge
ComfyUI Node

Vision LLM Judge

A ComfyUI node in Kinburg-Nodes/LLM with 10 inputs and 4 outputs.

By Kinburg·Created 2 months ago·Updated 3 days ago· 1
Vision LLM Judge
  • config
  • images
  • summary
  • results_json
  • best_index
  • help
rubricRate the overall image quality: anatomy/structure, sharpness and detail, and how well it matches the prompt. Penalize artifacts, extra or malformed limbs/fingers, and watermarks.
prompts
score_min1
score_max5
unload_after_runconfig default
criteriaoverall_quality: style matches the prompt, no artifacts, no excess noise, correct proportions, good color reproduction anatomy: all required limbs present, no extra limbs, correct placement, natural pose, proportional body prompt_compliance: how accurately the image follows the generation prompt camera: camera angle and camera settings match the intent text: if the prompt requests text — present, character-accurate, correct color/font/size/placement (if the prompt has NO text, give the top score)
system_promptYou are a strict, consistent image-quality judge. You are shown ONE image plus evaluation criteria. Assess the image only against those criteria and be objective. Reply with ONLY a single JSON object — no prose, no markdown, no code fences.
comment_styleone concise sentence
CategoryKinburg-Nodes/LLM

Inputs (10)

NameTypeDefaultDescription
configKINBURG_LLM_CONFIGA 'Local LLM Settings (GGUF)' node WITH a 'Vision Settings (GGUF)' (mmproj) attached. The judge needs vision.
imagesIMAGEImages to judge — a batch OR an image list (mixed sizes fine).
rubricSTRINGRate the overall image quality: anatomy/structure, sharpness and detail, and how well it matches the prompt. Penalize artifacts, extra or malformed limbs/fingers, and watermarks.What to judge each image on. Be concrete.
promptsoptSTRINGOptional per-image generation prompts, '---'-separated blocks (e.g. Get Accumulator (prompts)). Lets the judge assess prompt adherence.
score_minoptINT10–100Lowest score on the scale.
score_maxoptINT51–100Highest score on the scale.
unload_after_runoptCOMBOconfig defaultFree the model from VRAM after THIS node runs, without touching the shared config. 'config default' follows the Settings node; 'unload after run' frees VRAM before your image generation; 'keep loaded' stays warm.
criteriaoptSTRINGoverall_quality: style matches the prompt, no artifacts, no excess noise, correct proportions, good color reproduction anatomy: all required limbs present, no extra limbs, correct placement, natural pose, proportional body prompt_compliance: how accurately the image follows the generation prompt camera: camera angle and camera settings match the intent text: if the prompt requests text — present, character-accurate, correct color/font/size/placement (if the prompt has NO text, give the top score)Multi-criteria mode: one criterion per line, 'name' or 'name: what it means'. The judge scores EACH on the score scale and the overall score is their average. Pre-filled with an example — edit it, or CLEAR it for a single overall score.
system_promptoptSTRINGYou are a strict, consistent image-quality judge. You are shown ONE image plus evaluation criteria. Assess the image only against those criteria and be objective. Reply with ONLY a single JSON object — no prose, no markdown, no code fences.Who the judge is (the model's system prompt). The built-in default is shown — edit to change persona/strictness. Cleared → falls back to the built-in default.
comment_styleoptSTRINGone concise sentenceHow the 'comment' field should read, inserted into the JSON instruction. Default 'one concise sentence'; e.g. 'two to four sentences covering strengths and weaknesses' for a detailed review (raise the Settings max_tokens for long comments).

Outputs (4)

NameTypeDescription
summarySTRING
results_jsonSTRING
best_indexINT
helpSTRING