abenzerps Q8_0 ships 1D RMSNorms as packed Q8 (136 vs 128), which breaks
Comfy rms_rope; tagger now dequantizes small tensors and the graph uses
TextEncodeQwenImage21's 64-ch latent plus AuraFlow shift.
Co-authored-by: Cursor <cursoragent@cursor.com>
needsFit only compared aspect ratio, so a same-aspect-but-larger still
(e.g. a 4032x2268 phone photo against a 1344x768 canvas) would skip
fitting and hand MiniMaxH3AddGuide a still that does not match the
job's own latent size. Require an exact width+height match instead;
anything else (including unreadable dimensions) still fits.
Also guard the video save path: an empty Comfy download, an empty
stitched assembly, or a stitched result with no readable duration now
fails the job with a clear error instead of silently saving/attaching
a broken zero-length clip.
Co-authored-by: Cursor <cursoragent@cursor.com>
upload() ran fitStill (ffmpeg scale+pad) on every image except hero,
including the start/handoff still. Four chained extends resampled the
anchor frame four times and destroyed identity even with the hero
reference and lock text intact.
- upload(): only fit when the still's aspect truly differs from the
job canvas (needsFit); never fit hero or start regardless.
- Track handoffSource (anchor|extract) and whether the start frame was
fitted; surface both in job metadata and Details.
- extendSelected(): re-binding an already-selected clip on task switch
must not stomp the user's overridden locks/settings back to the
clip's saved values (regression introduced by the last hotfix push).
- Add tests/studio2-extend-fit.test.mjs: two extends at identical
width/height reach Comfy with the anchor PNG byte-identical, no
ffmpeg fit.
Co-authored-by: Cursor <cursoragent@cursor.com>
