Release Notes

This is a smaller version of the model, trained at a higher resolution using musubi-tuner. Based on my testing so far, this version has produced the best overall balance of motion quality, temporal stability, and fine detail among the versions I have trained.

The LoRA is designed to work well at a strength of 1.0, which is currently my recommended starting point. Lower strengths are also valid and can produce good results, depending on the base model, prompt, workflow, and the other LoRAs being used. In my testing, this strength generally improves the visibility of the learned details and motion characteristics, although the optimal value will depend heavily on the rest of the pipeline.

Please keep in mind that LoRA results are highly dependent on the generation environment. Different inference implementations, samplers, schedulers, quantization settings, model versions, resolutions, prompts, and combinations of other LoRAs can all affect the final result. Because of this, two people can use the same LoRA and obtain noticeably different results. There is no single configuration that will necessarily be optimal for everyone.

As always, this model is provided completely free of charge. Please feel free to use it, modify your workflow, experiment with different strengths, or simply choose whatever configuration works best for your particular setup.

For transparency and reproducibility, all of my example videos for my LoRAs have the workflow embedded, including the relevant prompts, LoRA settings, model configuration, and other generation parameters. If you are having difficulty reproducing the results, I strongly recommend downloading one of the example videos and loading its embedded workflow. This gives you a known-good reference configuration and makes troubleshooting significantly easier.

If the result still differs from the example, that does not necessarily indicate a problem with the LoRA. It may be worth comparing the complete environment and pipeline, including the base model/checkpoint, model revision, VAE, resolution, sampler/scheduler, seed, LoRA weights, inference software/version, precision or quantization settings, and any additional nodes or processing steps.

If you encounter a reproducible problem, please feel free to mention it in the comments. I read the feedback and take technical issues seriously. I may not always respond to every comment, but detailed and constructive feedback is always very welcome and helps me improve future versions.

Thank you to everyone who tests these models, reports their results, and shares feedback. I really appreciate it.