NAI style LoRA for MiniMax H3 (Ref2VA). The soft, bright NAI anime look in motion: light pastel colours, thin delicate line work and glossy shading.
Why it exists: plain image-to-video keeps an NAI still's look, but Ref2VA (a character sheet as the reference) drifts back to generic 2.5D anime. This LoRA was trained with reference pictures attached, so the look holds when you work from a sheet.
Trigger: none. Just load it.
Strength: 0.85 is the sweet spot. 0.7 gives a lighter touch. 1.0 is the full look, but it starts to redraw clothes (a black shirt turned white, a sash disappeared). Below 0.5 it does almost nothing.
Model: trained and tested on MiniMax H3 Ref2VA (reference-to-video). Not tested on FL2VA or plain I2V.
Speed LoRAs: works with the turbo / lightx2v LoRA at 10 steps.
With NAI-trained character LoRAs: don't stack it at 0.7 or above. The two carry the same look, and together they over-gloss the skin and rewrite the other characters. The character LoRA alone is usually enough.
Rank 16, 148 MB.
First release. Trained for MiniMax H3 Ref2VA with reference pictures, so the NAI look holds when you give it a character sheet.
No trigger word. The look is always on; strength is the dial.
Strength 0.85 for the NAI look with the reference's clothes kept. 1.0 lets the LoRA redraw outfits.
4,000 steps, rank 16, tested with the turbo LoRA at 10 steps.
Description
First release. Trained for MiniMax H3 Ref2VA with reference pictures, so the NAI look holds when you give it a character sheet.
No trigger word. The look is always on; strength is the dial.
Strength 0.85 for the NAI look with the reference's clothes kept. 1.0 lets the LoRA redraw outfits.
4,000 steps, rank 16, tested with the turbo LoRA at 10 steps.