Version 0.1.4

What's Changed

• feat: SkyReels V1 I2V training by @kohya-ss in #97

Full Changelog: v0.1.3...v0.1.4

Version 0.1.3

What's Changed

• Add python project config via uv while keeping backward compatibility with pip install method by @bmaltais in #51
• Fix some bugs of full fine turning. by @wangqiang9 in #40
• Update lycoris for hv_generate_video.py by @sdbds in #56
• Support Skyreels-I2V by @sdbds in #95
• Update for loading Safetensors vae by @sdbds in #78
• feat: add support for inference with SkyReels V1 T2V and I2V models by @kohya-ss in #96
• fix: rope None at MMSingleStreamBlock by @bruce2233 in #84
• fix bugs of datetime.timedelta by @wangqiang9 in #82
• Add exclude single blocks to hv_generate_video.py to only load double blocks by @maybleMyers in #69

New Contributors

• @bmaltais made their first contribution in #51
• @bruce2233 made their first contribution in #84
• @maybleMyers made their first contribution in #69

Full Changelog: v0.1.2...v0.1.3

Version 0.1.2

What's Changed

• Add LoRA Merger script, Updates README files. by @kaykyr in #37
• Set some trainable parameters of transformer for full training. by @wangqiang9 in #39
• V2v by @kohya-ss in #47
• feat: delete cache files not in dataset by @kohya-ss in #53

New Contributors

• @kaykyr made their first contribution in #37
• @wangqiang9 made their first contribution in #39

Full Changelog: v0.1.1...v0.1.2

v0.1.1

What's Changed

• Add support for sampling during training by @NSFW-API in #12 Refer to this document for details.
• Sample images during training by @kohya-ss in #28
• add lycoris support by @sdbds in #33
• Add option to specify fps in generate command by @nivibilla in #34
• You can now specify the number of repetitions for each dataset. The dataset is repeated the specified number of times, and the training is performed as one epoch. Specify num_repeats in the .toml. For details, please refer to this document.
• LoRA now excludes img_mod and txt_mod of double blocks and modulation of single blocks by default. According to reports from the community, this has improved the training results. You can change the target modules by specifying exclude_patterns and include_patterns with --network_args. For details, please refer to this document.
  • If you are resuming training by specifying the previous weights with --network_weights, please specify --network_args "include_patterns=[r'.*(img_mod|txt_mod|modulation).*']".
• LoRA+ is now available. Specify loraplus_lr_ratio in --network_args. For details, please refer to this document.
• Added a temporary --save_merged_model option to hv_generate_video.py to save the DiT model after LoRA merge.

New Contributors

• @NSFW-API made their first contribution in #12
• @nivibilla made their first contribution in #34

Full Changelog: v0.1.0...v0.1.1

Pre-release v0.1.0

This is the first pre-release of Musubi Tuner. Basic functionality has been implemented and tested. While under active development, this version provides a stable reference point for documentation writers and third-party tool developers.

Key features:

• Caching latentss and text encoder output for training.
• Training HunyuanVideo LoRA.
• Inference HunyuanVideo with LoRA.