How guandeh17/self-forcing works
Self Forcing is a research system for autoregressive video diffusion, with GUI and CLI generation as its user-facing path and training methods as a second major capability. Inference encodes prompts, performs causal or bidirectional diffusion sampling, then decodes video; causal sampling maintains KV caches to support chunked generation. The repository also includes Wan text-to-video and image-to-video runtimes, training objectives, and data preparation utilities. Some training components are identified from the tree and README only; their internal calls are not asserted.
Main entry points: wan_wrapper.py Wan wrappers · trainer Training loops · model.py Wan transformer · causal_diffusion_inference.py Causal sampling
Components 30
User entry points
- GUI demo
demo.py - CLI inference
inference.py - Training command
train.py
Generation pipelines
- Causal sampling
pipeline/causal_diffusion_inference.py - Causal inference
pipeline/causal_inference.py - Bidirectional sampling
pipeline/bidirectional_diffusion_inference.py - Bidirectional inference
pipeline/bidirectional_inference.py - Self-forcing rollout
pipeline/self_forcing_training.py - Flow solvers
wan/utils/fm_solvers.py - UniPC solver
wan/utils/fm_solvers_unipc.py
Wan model runtime
- Wan wrappers
utils/wan_wrapper.py - Text encoder
wan/modules/t5.py - Tokenizers
wan/modules/tokenizers.py - Wan transformer
wan/modules/model.py - Causal transformer
wan/modules/causal_model.py - Attention kernels
wan/modules/attention.py - Video VAE
wan/modules/vae.py - Wan text-to-video
wan/text2video.py - Wan image-to-video
wan/image2video.py - Context parallelism
wan/distributed/xdit_context_parallel.py
Training methods
- DMD objective
model/dmd.py - CausVid objective
model/causvid.py - GAN objective
model/gan.py - SID objective
model/sid.py - ODE regression
model/ode_regression.py - Training loops
trainer
Training data
- Prompt dataset
utils/dataset.py - LMDB access
utils/lmdb.py
Other components
- User
- External checkpoints
Connections 36
- User → GUI demo: uses
README.md - User → CLI inference: runs
README.md - GUI demo → Wan wrappers: loads models
demo.py - GUI demo → Causal sampling: uses pipeline
- CLI inference → Causal sampling: selects pipeline
- CLI inference → Prompt dataset: reads prompts
inference.py - Causal sampling → Wan wrappers: calls
pipeline/causal_diffusion_inference.py - Causal sampling → Flow solvers: schedules steps
pipeline/causal_diffusion_inference.py - Causal sampling → UniPC solver: uses solver
pipeline/causal_diffusion_inference.py - Causal inference → Wan wrappers: calls
pipeline/causal_inference.py - Bidirectional sampling → Wan wrappers: calls
pipeline/bidirectional_diffusion_inference.py - Bidirectional sampling → Flow solvers: schedules steps
pipeline/bidirectional_diffusion_inference.py - Bidirectional sampling → UniPC solver: uses solver
pipeline/bidirectional_diffusion_inference.py - Bidirectional inference → Wan wrappers: uses models
pipeline/bidirectional_inference.py - Wan wrappers → Text encoder: loads encoder
utils/wan_wrapper.py - Wan wrappers → Tokenizers: uses tokenizer
utils/wan_wrapper.py - Wan wrappers → Wan transformer: loads model
utils/wan_wrapper.py - Wan wrappers → Causal transformer: loads causal model
utils/wan_wrapper.py - Wan wrappers → Video VAE: loads decoder
utils/wan_wrapper.py - Causal transformer → Wan transformer: reuses layers
wan/modules/causal_model.py - Causal transformer → Attention kernels: uses attention
wan/modules/causal_model.py - Wan transformer → Attention kernels: uses attention
wan/modules/model.py - Wan text-to-video → Wan transformer: runs model
wan/text2video.py - Wan image-to-video → Wan transformer: runs model
wan/image2video.py - Wan image-to-video → Context parallelism: enables parallelism
wan/image2video.py - Context parallelism → Wan transformer: patches forward
wan/distributed/xdit_context_parallel.py - Training command → Training loops: starts training
- Training loops → Self-forcing rollout: uses rollout
- Training loops → Prompt dataset: loads prompts
trainer/distillation.py - Prompt dataset → LMDB access: reads records
utils/dataset.py - Training loops → DMD objective: trains objective
- Training loops → CausVid objective: trains objective
- Training loops → GAN objective: trains objective
- Training loops → SID objective: trains objective
- Training loops → ODE regression: trains objective
- External checkpoints → Wan wrappers: provides weights