[vikasprogrammer]

Why

salonking-social's Wan 3.0 / LTX tests were blocked by video_generate, not by the models:

• alibaba/wan-3.0/reference-to-video + image → an unrelated velvet chair (the model reads refers, a list of {url,type}; image was dropped in silence). The same gap applies to Seedance 2.0 mini (reference_images) and Kling O3 (images).
• ltx-2.3-quality/image-to-video → 400 image_size_obj (its seed field is image_url, and pricing needs resolution); ltx/ltx-2.3-fast/image-to-video → 400 resolution is required.

What

• The Atlas adapter reads the model's published Input schema (/api/v1/models → schema URL, cached 6h; falls back to the old field names) and maps the seed, referenceImages, the duration (snapped to the enum, or num_frames = 8k+1 @24fps) and aspectRatio (a ratio field, or a resolution preset like portrait_9_16) onto it.
• It fills unset enum fields with their schema defaults.
• An input the model has nowhere to put is now an error naming the accepted fields, not a silent drop.
• Inline images going to a non-image field are hosted through Atlas uploadMedia first, because Wan refers requires a public URL.
• params is passed verbatim and wins. Unknown keys still go out, with a note in the tool result.
• fal: aspect_ratio, image_urls and params are passed through.

Verified

• scripts/video-gen-mapping-test.cjs (added to test:governance), plus the full test:governance run: green.
• Live on Atlas with the salonking still: Wan 3.0 reference-to-video at 9:16 gave 720×1280, 5s, showing the input chair. LTX 2.3 quality i2v at 9:16 gave 576×1024, 5s, no 400.

🤖 Generated with Claude Code