What problem does it solve? AI image and video generation pipelines produce inconsistent output, and manually reviewing every result does not scale. This Skill automates quality gating by having a vision model (MLLM) act as a judge, scoring outputs against a rubric instead of relying on rigid metrics like FVD or CLIP that diverge from human perception. ## Core Features & Use Cases - Cheap Pre-Filter: Rejects obviously broken outputs (black frames, blur, zero-byte files, duplicates, frozen video) with lightweight JavaScript checks before spending any vision-model calls. - Rubric-Based Scoring: Scores each output 0-5 on prompt adherence, motion and temporal consistency, aesthetics, and physical artifacts, with a one-line justification per axis. - Verdict and Best-of-N: Produces ACCEPT/REJECT decisions against a configurable threshold, suggests one concrete fix for regeneration, and ranks multiple candidates to pick the best. - Use Case: In a batch pipeline generating 30 product images, automatically reject the blurry or off-prompt ones, get a specific prompt fix for each rejection, and only pass clean results to download. ## Quick Start Evaluate this generated image against the brief using the mllm-judge rubric and give me an accept or reject verdict with scores.
