ComfyUI Node Runs on cloud

Judge Images

Score each picture in a batch, as several seeds of one prompt make, from 0 to 10 with a language model that reads them. Edits are scored on following, consistency and quality against their source; generations on following and quality. Hands on the best and its notes.

By WASasquatch·Created 4 years ago·Updated 3 days ago· 1,900
Judge Images
  • clip
  • images
  • source
  • best
  • index
  • passed
  • scores
  • notes
◄prompt►
◄least_following9►
◄least_consistency8►
◄least_quality6►
CategoryWAS Suite/Image/Analyze

Inputs (7)

NameTypeDefaultDescription
clipCLIPqwen3.5_9b_int8_convrot.safetensors from Load CLIP with type krea2, or another Qwen-VL model.
imagesIMAGEThe pictures to score, 1 to 6, as VAE Decode gives a batch from Repeat Latent Batch.
promptSTRINGThe instruction the edits followed or the prompt the pictures came from, as 'Make it snow'.
least_followingINT90–10The following score a picture needs to pass, 0 to 10.
least_consistencyINT80–10The consistency score an edit needs to pass, 0 to 10. Unused without a source.
least_qualityINT60–10The quality score a picture needs to pass, 0 to 10.
sourceoptIMAGEThe picture before the edit. Unwired = the pictures are judged as generations.

Outputs (5)

NameTypeDescription
bestIMAGEThe picture kept: the best that passes, or the best of all when none does.
indexINTThe kept picture's place in the batch, from 0.
passedBOOLEANTrue when the kept picture reaches every least score.
scoresSTRINGEvery picture's scores and notes as a JSON list, in batch order.
notesSTRINGThe flaws found in the kept picture, for Generate Edit Instruction's feedback.