CRITICAL FIX: WAN models require binary masks (0.0 or 1.0), not gradient masks
- Replace gradient mask logic with binary masks throughout
- mask=0.0: Use provided keyframe image (inpainting guidance)
- mask=1.0: Generate freely (no guidance)
- Context window pyramid fusion handles smooth blending automatically
- Gradient masks were causing ghosting, blurriness, and artifacts
Technical details:
- Start keyframe: mask=0.0 for first frame_blend_width frames
- Middle keyframe: mask=0.0 for frame_blend_width region centered at middle_frame_position
- End keyframe: mask=0.0 for last frame_blend_width frames
- All other regions: mask=1.0 (fully generated)
- Context windows with pyramid fusion create smooth transitions at overlaps
Fixes tensor dimension error and improves visual quality
Bump version to 1.2.1
- Replace hard mask edges with smooth gradient transitions
- Start frame: mask gradient from 0.0 (preserve) to 1.0 (generate)
- Middle frame: V-shaped mask (1.0 → 0.0 at center → 1.0)
- End frame: mask gradient from 1.0 (generate) to 0.0 (preserve)
- This creates natural blending between keyframes and generated content
- Prevents hard cuts and improves visual coherence
- Example: with frame_blend_width=22
- Frame 0: 100% keyframe preserved
- Frame 11: 50/50 blend of keyframe and generation
- Frame 21: 100% AI generated content
- Keyframes are now repeated across the entire blend_width region
- Previously only unmasked as many frames as provided in input images
- Now single-frame keyframes are extended to fill the blend region
- Start frame: repeated from frame 0 to frame_blend_width
- Middle frame: repeated across frame_blend_width region centered at middle_frame_position
- End frame: repeated across last frame_blend_width frames
- This provides much stronger anchoring and reduces blurry inpainted regions
- Fixes issue where most of video was masked with single-frame keyframes
- Fix middle frame mask to only unmask regions with actual image data
- Fix end frame mask to match actual copied frame length
- Previously was unmasking much larger regions filled with gray (0.5)
- This was causing the model to preserve gray as 'real' data instead of inpainting
- Now only unmasks middle_img_len and end_img_len frames with +3 buffer for latent encoding
- Resolves issue where middle section of generated video turns gray
- Implement WanThreeFrameToVideo node for generating videos from 3 keyframes (start, middle, end)
- Add adjustable middle frame positioning (0.0-1.0 for flexible timing)
- Add configurable frame blend width for smooth keyframe transitions
- Support CLIP vision output concatenation for enhanced conditioning
- Include monkey patch to fix context window bug in ComfyUI for WAN models
- Fixes tensor subsetting in model_conds dictionary (concat_latent_image, concat_mask)
- Enables proper context window support for all WAN nodes
- Integrate node into existing custom node pack structure
- Update README and pyproject.toml with new node documentation
- Bump version to 1.2.0
MariusKM