Upload 6 files
s3_fl2va_calibrated.safetensors is an extracted lora V2 then calibrated using clips with matching prompts from training dataset in pairs of 2 each time : two showing concepts in dataset wanted to retain and two showing general video/audio quality generated by base model to preserve. at every round One clip from each category guided the tuning; the other two checked the results. The tune code encoded the clips, added noise, and tested denoising predictions to adjust LoRA strengths without generating full videos for every layer. The aim was to balance concept retention and quality, with results checked only on those four examples.
S3_fl2va_lora.safetensors is an extracted lora V1 nothing new just big rank extraction
minimax_S3_tuned_Tame.safetensors is an extracted lora V1(near full s3 extracted training) big rank , tuned by hand layer per layer for best results with added rank i think to it ... that why too big
minimax_h3_max.safetensors is extracted lora V2 (s3_fl2va_calibrated.safetensors) full power tuned by hand layer per layer for best results
minimax_S3_tuned_TameV2.safetensors is extracted lora V2 (s3_fl2va_calibrated.safetensors) tamed power tuned by hand layer per layer for best results
minimax_S3_tuned_normal.safetensors is extracted lora V2 (s3_fl2va_calibrated.safetensors) med power tuned by hand layer per layer for best results
the best one might differ on what you need , if you want to avoid baking the model minimax_S3_tuned_Tame.safetensors is my favorite ! next to it and smaller in size minimax_S3_tuned_TameV2 minimax_S3_tuned_normal are really good
extracted lora V2 and extracted lora V1 differ in logic
all loras mute or lower audio layers so you will find audio less distorted and more close to base model