Tune visual parameters with vision-model feedback loops

An automated visual eval loop that uses a vision or video-understanding model as the judge for tuning parameters where "looks/feels right" is the success criterion and no cheap numeric metric exists - animation easing and timing, camera or zoom feel, color grade, layout and spacing, design params, render or encoder settings, or prompt params. It is the automated counterpart to human lookdev review, for use when no human is available to sit the loop. Its worked reference is the screenstudio-alternative skill's iteration.py, which tuned zoom-animation feel via fal-ai/video-understanding. The loop has four steps: render N labeled variants into one artifact with the parameter values burned directly onto the image or video - a labeled grid for images, a labeled sequence with overlay cards for video so the model can compare temporally; make one model call with an explicit rubric defining what "good", "too much", and "too little" look like, asking for structured JSON output with per-variant ratings and concrete suggested new values, for example {"ratings":{"A":n,...},"best_so_far":"X","suggest":[[p1,p2],...]}; go coarse to fine, with round 1 as a wide spread to locate the region and round 2 rendering the model's suggestions plus the current best into one artifact for a single pick, usually converging in two rounds; and stop once the best rates high and suggestions cluster. It documents specific cost and quality reduction techniques: one artifact per round rather than one call per variant, burning params onto the artifact itself, structured JSON output parsed by regex rather than free-text wrangling, tuning on a short representative sample rather than the full asset, capping variants at 5-6, including a deliberately-bad and a safe-default calibration anchor each round, stating an independent rubric up front rather than asking "which do you like" since that lets the model echo the user's framing, reusing renders across rounds instead of re-rendering the winner, an early-exit rule when round 1's top score is 9+/10 with clustered suggestions, and picking the cheapest judge that can actually see the failure - an image VLM for spatial judgments, a true video model only for temporal ones like easing or motion smoothness that are invisible in stills.

Use it when the target is a "does this look/feel right" judgment with no numeric metric to optimize directly. Do not use it when a real numeric metric exists and correlates with quality - optimize that directly instead of paying a model per step; when the judgment is genuinely subjective to the specific user's taste or brand, since a model's "best" isn't necessarily their best (the source notes this is exactly why a screen-studio spring auto-tune was dropped, because the model's pick didn't match the owner's eye); or when there are only one or two variants, where looking yourself is simpler.

Input is a small set of rendered variants - images or video clips - with their parameter values, plus an explicit rubric for what "good" means. Output is structured JSON with per-variant ratings and suggested new parameter values, following the documented schema:

{"ratings":{"A":n,...},"best_so_far":"X","suggest":[[p1,p2],...]}

Built around vision-language models for spatial or image judgments and video-understanding models like fal-ai/video-understanding for temporal judgments such as easing, timing, and motion smoothness; references the screenstudio-alternative skill's iteration.py as a worked implementation, and the source's own verify-outputs rule for keeping the judging rubric independent of what was tuned.

Engineers and coding agents tuning visual or motion parameters - animation, color grade, layout, render settings - who need an automated judge in place of a human sitting the loop, but who still sanity-check the model's pick against a safe default before committing.