How yaofang-liu/pusa-vidgen works
Pusa is a video-diffusion system centered on frame-level vectorized timestep adaptation, with Wan2.1/Wan2.2 support and text-to-video, image-to-video, and video-to-video workflows. The principal path runs from a task input through a task-specific pipeline to generated video; the repository also includes DiffSynth model loading, model components, and optional processing. The graph combines README-documented capabilities with relationships visible in indexed source references; several pipeline internals are unsampled, so uncertain connections are omitted.
Main entry points: pipelines_ti2v_release.py Text/Image Video · asymm_models_joint.py Joint DiT · pipeline_runner.py Pipeline Runner · model_manager.py Model Manager
Components 24
Pusa Pipelines
- Text/Image Video
src/genmo/pusa/pipelines_ti2v_release.py - Multi-Frame Tasks
src/genmo/pusa/pipelines_multi_frames_release.py - Video-to-Video
src/genmo/pusa/pipelines_v2v_release.py
DiffSynth Pipelines
- Pipeline Runner
PusaV1/diffsynth/pipelines/pipeline_runner.py - Wan Pusa Video
PusaV1/diffsynth/pipelines/wan_video_pusa.py - Wan2.2 Pusa
PusaV1/diffsynth/pipelines/wan22_video_pusa.py
Model Runtime
- Model Manager
PusaV1/diffsynth/models/model_manager.py - Model Registry
PusaV1/diffsynth/configs/model_config.py - Wan DiT
PusaV1/diffsynth/models/wan_video_dit.py - Wan Pusa Model
PusaV1/diffsynth/models/wan_video_pusa.py - Wan Video VAE
PusaV1/diffsynth/models/wan_video_vae.py
Pusa Core
- Joint DiT
src/genmo/pusa/dit/joint_model/asymm_models_joint.py - DiT Layers
src/genmo/pusa/dit/joint_model/layers.py - Context Parallelism
src/genmo/pusa/dit/joint_model/context_parallel.py - Pusa VAE
src/genmo/pusa/vae/models.py - Parallel VAE Convs
src/genmo/pusa/vae/cp_conv.py
Data Processing
- Embedding Datasets
src/genmo/pusa/datasets/embedding_datasets.py - Video Transforms
src/genmo/pusa/datasets/video_transforms.py - Video Data
PusaV1/diffsynth/data/video.py - Video Processors
PusaV1/diffsynth/processors/sequencial_processor.py
Other components
- Text Prompt
- Input Image
- Input Video
- Generated Video
Connections 27
- Text Prompt → Text/Image Video: prompts
README.md - Input Image → Text/Image Video: conditions
README.md - Input Image → Multi-Frame Tasks: conditions
README.md - Input Video → Video-to-Video: conditions
README.md - Text/Image Video → Generated Video: generates
README.md - Multi-Frame Tasks → Generated Video: generates
README.md - Video-to-Video → Generated Video: generates
README.md - Text/Image Video → Joint DiT: runs DiT
- Multi-Frame Tasks → Joint DiT: runs DiT
- Video-to-Video → Joint DiT: runs DiT
- Joint DiT → Pusa VAE: decodes latents
- Text/Image Video → Embedding Datasets: uses embeddings
- Text/Image Video → Video Transforms: uses transforms
- Joint DiT → DiT Layers: uses layers
- Joint DiT → Context Parallelism: uses context
- Pusa VAE → Parallel VAE Convs: uses convolutions
- Text Prompt → Pipeline Runner: submits prompt
- Pipeline Runner → Wan Pusa Video: dispatches
- Pipeline Runner → Video Data: uses data
PusaV1/diffsynth/pipelines/pipeline_runner.py - Pipeline Runner → Video Processors: uses processors
PusaV1/diffsynth/pipelines/pipeline_runner.py - Pipeline Runner → Generated Video: returns video
- Wan Pusa Video → Wan Pusa Model: runs model
- Wan Pusa Video → Wan Video VAE: uses VAE
- Wan2.2 Pusa → Wan Pusa Model: runs model
- Model Manager → Model Registry: loads registry
PusaV1/diffsynth/models/model_manager.py - Model Manager → Wan DiT: loads model
PusaV1/diffsynth/models/model_manager.py - Model Manager → Wan Pusa Model: loads model
PusaV1/diffsynth/models/model_manager.py