How vita-epfl/stable-video-infinity works
Stable Video Infinity (SVI) is a long-video generation system built on DiffSynth. The intended user-to-result path is prompt and conditioning input, video generation, then a rendered long-form video; README highlights extending video length while recycling generation error. The sampled Wan pipeline shows model, scheduler, prompt, and memory-management dependencies, but the SVI-specific pipeline and UI implementations were not sampled. Dashed, uncited links mark those uncertain runtime connections; the README excerpt does not document detailed implementation flow.
Main entry points: wan_video.py Wan video pipeline · svi_video.py Long video pipeline · gradio_demo.py Demo interface · model_manager.py Model manager
Components 16
User entry
- Demo interface
gradio_demo.py
SVI generation
- Long video pipeline
diffsynth/pipelines/svi_video.py - Error recycling
diffsynth/pipelines/svi_video.py
DiffSynth runtime
- Wan video pipeline
diffsynth/pipelines/wan_video.py - Model manager
diffsynth/models/model_manager.py - Wan transformer
diffsynth/models/wan_video_dit.py - Wan text encoder
diffsynth/models/wan_video_text_encoder.py - Wan video VAE
diffsynth/models/wan_video_vae.py - Wan image encoder
diffsynth/models/wan_video_image_encoder.py - Flow scheduler
diffsynth/schedulers/flow_match.py
Conditioning
- Wan prompter
diffsynth/prompters/wan_prompter.py - Pose extraction
utils/run_align_pose.py - Audio processing
utils/audio_process.py - Video data
diffsynth/data/video.py
Other components
- Video creator
- Generated video
Connections 15
- Video creator → Demo interface: submits request
- Demo interface → Long video pipeline: starts generation
- Long video pipeline → Error recycling: recycles error
- Long video pipeline → Wan video pipeline: uses generation
- Wan video pipeline → Wan transformer: configures model
diffsynth/pipelines/wan_video.py - Wan video pipeline → Wan text encoder: configures encoder
diffsynth/pipelines/wan_video.py - Wan video pipeline → Wan video VAE: configures VAE
diffsynth/pipelines/wan_video.py - Wan video pipeline → Wan image encoder: configures encoder
diffsynth/pipelines/wan_video.py - Wan video pipeline → Flow scheduler: uses scheduler
diffsynth/pipelines/wan_video.py - Wan video pipeline → Wan prompter: creates prompter
diffsynth/pipelines/wan_video.py - Wan video pipeline → Model manager: imports manager
diffsynth/pipelines/wan_video.py - Long video pipeline → Pose extraction: uses pose input
- Long video pipeline → Audio processing: uses audio input
- Long video pipeline → Video data: handles video
- Long video pipeline → Generated video: produces video