Skip to content

Changelog

All notable changes to Smart-Diffusion will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.1.2] - 2026-03-16

Changed

  • Unified launch entry to run.sh only
  • Moved startup/system parameter control to system_config.yaml
  • Replaced CFG parallel toggle with explicit cfp -> infer.diffusion.cfg_size flow
  • Updated documentation to use bash run.sh system_config.yaml ... examples
  • Extended eval strategy support to multi-select eval.eval_type=[...]
  • Added eval strategies: fid, fvd, psnr, ssim, lpips (in addition to vbench)
  • Added eval.reference_path for reference-based metrics, with skip-on-missing behavior

[0.1.1] - 2026-02-16 Chuxi

Added

  • Comprehensive English docstrings for core modules
  • Enhanced README.md with:
  • Improved structure and formatting
  • Comprehensive feature descriptions
  • Detailed installation instructions with uv support
  • Usage examples and configuration guide
  • Contributing guidelines and roadmap
  • Complete documentation website using MkDocs Material
  • Installation guide
  • Quick start tutorial
  • Architecture overview
  • FAQ section
  • Configuration guide
  • GitHub Actions workflow for automatic documentation deployment
  • Support for documentation search and code highlighting

Changed

  • Reorganized documentation structure with clear sections
  • Improved code documentation standards across the codebase

[0.1.0] - 2026-01-27

Added

  • Initial release of Smart-Diffusion
  • Support for Wan-T2V series models (1.3B, 14B, A14B)
  • Multiple attention backend support:
  • FlashAttention (default)
  • SageAttention (quantized)
  • SpargeAttention (sparse)
  • Memory optimization features:
  • Low memory mode with model offloading
  • VAE tiling support
  • Multi-level memory management (0-3)
  • FlexCache system for feature reuse:
  • TeaCache strategy
  • Pyramid Attention Broadcast (PAB) strategy
  • Parallelism support:
  • Context Parallelism (CP)
  • Classifier-Free Guidance (CFG) parallelism
  • Evaluation support:
  • VBench custom-mode evaluation
  • Configuration system:
  • Hydra-based configuration
  • Three-tier parameter system (Model/User/System)
  • Task management:
  • Task pool and scheduler
  • Request serialization for distributed execution

Known Issues

  • Data parallelism not yet implemented
  • Limited model support (Wan-T2V only)
  • Documentation incomplete in some areas

Future Roadmap

Planned Features

  • [ ] Models
  • [ ] Flux-2
  • [ ] FireRed-Image-edit
  • [ ] Longcat
  • [ ] AutoVideoParallel
  • [ ] DiTango
  • [ ] Hybrid parallelism combinations
  • [ ] More acceleration algorithms
  • [ ] Additional cache strategies
  • [ ] Quantization improvements
  • [ ] Production features
  • [ ] HTTP API server
  • [ ] Batching and request queuing
  • [ ] Monitoring and metrics
  • [ ] Better operator implementations
  • [ ] Custom CUDA kernels
  • [ ] Triton implementations
  • [ ] Comprehensive benchmarks
  • [ ] Performance comparisons
  • [ ] Quality metrics

Documentation Improvements

  • [ ] Complete API reference for all modules
  • [ ] More usage examples
  • [ ] Video tutorials
  • [ ] Community contributions guide

Contributing

See Contributing Guide for how to contribute to Smart-Diffusion.

For detailed commit history, see GitHub Commits.