<img width="1219" height="790"
alt="{A15AF77F-C555-4C85-BC15-B4B8D44D2248}"
src="https://github.com/user-attachments/assets/ec4b52a2-68c4-42b0-88fc-00ee3b322321"
/>
A "recreate" button in the media lightbox modal.
On clicked, it will prefill the image or video promptbox with the same
model, prompt and ref images that was used to generate that media the
button was clicked from.
Temporary and doesnt support seedance video/audio ref yet.
- move a lot of "deprecated" crates so we don't use them
- kill several dead crates
- rename crates to be more semantic/clear
- roll back broken sha3 version bump
Paint/Mask cursor now draws properly regardless of canvas scale and drag
offset.
Inpainting now has its own brush with add and subtract actions.
A total node array is used for ordered-rendering of nodes, which fixes
the layer forward/backward operations.
3D nodes and drawings on the canvas were missing from the reference
image sent to the model in the default generation path. The fix mirrors
the Nano Banana Pro approach: render a composite of the base image and
the draw layer, upload it via uploadSceneSnapshot, and pass it as
scene_image_media_token rather than putting editedImageToken in
image_media_tokens.
Possible problem in our Seedance 2.0 provider means we need a backup.
This is here in case we need it.
Muapi is inferior for several reasons:
- Videos contain watermarks
- Videos cost 2x as much
- Not as many aspect ratio controls
- Not as many duration controls
- No starting frame / ending frame
- No video references
- No audio references
It does seem fast, however.
Generous use of caching used for react elements. Made a pretty
significant difference too.
Minor refactor to move helper functions outside instead of within
component tree.
BFlat
Brandon Thomas
Tanish Manku