- Added text_encoder/model-00004-of-00005.safetensors (4.7GB)
- Added text_encoder/model-00005-of-00005.safetensors (2.7GB)
- Total size: ~7.4GB
This completes the upload of all 36GB of One-to-All-Animation files!
Add One-to-All-Animation: Batch 1 - Small files and directories
FangSen9000+1: Claude
0a2e9f2
- Added LICENSE, README.md, requirements.txt
- Added datasets/, benchmark/, video-generation/, examples/, assets/
- Total size: ~100MB
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add the modified openpose and add confidence level filtering.
FangSen9000
adbfc5d
add more bug report
FangSen9000
cb41553
report a bug
FangSen9000
9dfadb7
Add verified posture setting
FangSen9000
2066a63
Add a setting that enables hiding and showing the main line.
FangSen9000
2a1ef7e
Update the demo of the smooth version of the pose video. The bottom part of the upper body needs to be optimized.
FangSen9000
fdc5154
This enables some controls to be turned on and off.
FangSen9000
78f9663
Discard the npz frames that cause black screen during generation
FangSen9000
0ebe26e
Restore the gallery
FangSen9000
9a3b4f8
Add a visual gallery demonstration
FangSen9000
1bbfe4e
Official version v0.1
FangSen9000
0c05e55
Achieve the interpolation effect at the npz level
FangSen9000
46a10a9
Fix minor errors in folder naming
FangSen9000
a36bbc9
Add optimized ControlNeXt-SVD-v2 implementation with direct pose input support
FangSen9000+1: Claude
6b0f4c6
Major improvements:
- Add run_controlnext_optimized.py with FP16 support, xformers, and GEGLU optimization
- Add direct pose video/images input to bypass DWPose extraction
- Implement smart output naming based on input filename
- Add ComfyUI-ControlNeXt-SVD reference implementation for analysis
- Update README with optimized inference examples
- Add easy_dwpose integration for pose extraction
- Generate example pose videos with text2pose system
Performance optimizations:
- Default FP16 precision for faster inference
- Memory-efficient attention with xformers
- Chunked VAE decoding to reduce memory usage
- GEGLU inplace operations for memory savings
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add ControlNeXt-SVD-v2 integration and optimizations
FangSen9000+1: Claude
2f8d0cb
- Add complete ControlNeXt-SVD-v2 model with pretrained weights
- Create optimized run_controlnext_fast.py with fp16 support and GEGLU patch
- Add conda environment configurations for controlnext
- Update README with optimized inference parameters
- Include generated video examples
- Add reference images for video generation
- Ensure all JPEG files are tracked with Git LFS
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Refactor to use individual NPZ files instead of merged database
FangSen9000+1: Claude
ac71c07
Major changes:
- Replace merged database approach with individual NPZ files per video
- Add WLASL_train.json mapping file and 9575 individual NPZ files in pose_dict/
- Create new text2pose_individual.py for working with individual NPZ files
- Add generated_pose_video/ directory for output videos
- Remove old pipeline scripts that worked with merged database
- Update file structure to be more modular and efficient
Benefits:
- More flexible video selection
- Easier to add/remove individual videos
- Better memory efficiency (load only needed videos)
- Cleaner file organization
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Update StableSigner3 repository structure and add Git LFS tracking