V5 - REF2VA
This version was refined from the FL2VA using only MLP from 10-49 to reduce interference with other loras and references, it has the same audio as FL2VA and was retrained to gain reference capabilities that FL2VA don't have, for this a dataset with references and goals is used. It was made with Musubi Trainer.
V5 - FL2VA
This version was trained with audio using the full layers spectrum, including both MLP and attention layers, with Musubi Trainer.
Compared to V4, it delivers more natural skin rendering and lighting, improved movement, more realistic results, and better audio quality. The dataset was increased by approximately 50%, and the training resolution was raised to capture more detail. Anatomy is also slightly improved, with better prompt adherence in some cases.
I tested it against V4 across more than 100 videos, and this version produced better results in approximately 70% of them. The validation results also improved, with a particularly significant improvement in T2VA performance.
V5 is based on FL2VA and uses all layers, resulting in a larger LoRA with a greater potential for interference with other LoRAs, particularly those targeting overlapping layers. If you encounter compatibility issues, let me know. A version trained without attention and identity layers may be a viable alternative, although removing these layers would likely reduce the overall quality.
New ref2va file to download (experimental, no audio trained)
Mystic XXX – Unlock Real Anatomy in T2V
Finally, text-to-video with proper anatomy. Works great with T2V, I2V and FFLF.
All example videos come with the full workflow embedded. Recommended strength: 0.2 – 1 (V4 I run at 1). Turbo LoRAs can change the look, adjust strength as needed.
Download, try it, and have fun!
Description
V5 - REF2VA
This version was refined from the FL2VA using only MLP from 10-49 to reduce interference with other loras and references, it has the same audio as FL2VA and was retrained to gain reference capabilities that FL2VA don't have, for this a dataset with references and goals is used. It was made with Musubi Trainer.
V5 - FL2VA
This version was trained with audio using the full layers spectrum, including both MLP and attention layers, with Musubi Trainer.
Compared to V4, it delivers more natural skin rendering and lighting, improved movement, more realistic results, and better audio quality. The dataset was increased by approximately 50%, and the training resolution was raised to capture more detail. Anatomy is also slightly improved, with better prompt adherence in some cases.
I tested it against V4 across more than 100 videos, and this version produced better results in approximately 70% of them. The validation results also improved, with a particularly significant improvement in T2VA performance.
V5 is based on FL2VA and uses all layers, resulting in a larger LoRA with a greater potential for interference with other LoRAs, particularly those targeting overlapping layers. If you encounter compatibility issues, let me know. A version trained without attention and identity layers may be a viable alternative, although removing these layers would likely reduce the overall quality.