rzgar/minimax_h3_fl2va_fp8_4steps

MiniMax H3 FL2VA 4-Step Merged Model + Attention/MLP LoRA Splits

The 4-step LoRA released by LightX2V | converted by KJ for MiniMax H3 FL2VA is a preview release and subject to change. This repo shares what I found while testing it, pand the artifacts built from it.

Clip 1 & 2: [4steps, er_sde/beta] | Clip 3: [6steps, dpmpp_2m/sgm_uniform] | Clip 4: [8steps er_sde/beta]

Facial/body morphing in wide shots appears to be a base-model issue not the LoRA

Multilingual impact

• Separated LightX2V LoRAs

During testing I noticed the 4-step LoRA can affect the multilingual capability of the base model. Based on a few generations in other languages (Persian, Kurdish), noticed a slight-to-heavy shift in pronunciation and broken sound in various words compared to the base model.

I separated the LoRA into its attention and MLP halves to isolate where the audio effect comes from:

• The MLP has the most obvious overall effect on audio compared to attention.
• Adjusting attention to 0.34 and MLP to 0.83 strength solved the audio issue but affected English negatively at the same strengths.

The split LoRAs are shared so you can experiment with the strengths yourself.

2-step preview mode

Increasing MLP to 1.34 (and audio shift to 1.73) can generate usable video (low motion) at 2 steps - very useful when you need lots of previews and prompt testing before investing in full-quality, high-step generations.

File	Download
minimax_h3_fl2va_fp16attn_fp8_4steps.safetensors	Download 28.2 GB