Minimax H3 character LoRA training with Ostris AI Toolkit

This tutorial covers character LoRA training for Minimax H3 in Ostris AI Toolkit, followed by use in ComfyUI. The presenter explains that images alone can teach appearance, while their voice workflow uses video clips with audio. They did not find a way to attach audio directly to images.

The example uses short 512 × 512 clips, mostly close-up images and manual video captions. Captions describe the character, setting and speech. The presenter recommends naming background sounds so the model does not associate them with the character. Image auto-captioning sometimes leaves captions unfinished or skips images, requiring another pass.

Training guidance includes disabling automatic frame counts for images and setting one frame, enabling automatic frame counts for videos, and caching latents for both datasets. The presenter reports out-of-memory errors with unsuitable frame settings and notes that system RAM matters alongside VRAM.

On a 5090 running Linux, the presenter reports roughly 4,000 seconds for 3,000 training steps. In their tests, voice resemblance emerged around 1,200 steps and visual likeness around 2,400. They caution that a reference LoRA can become too strong at 3,000 steps and suggest earlier checkpoints or gentler settings. A control prompt without a reference image helps assess what the LoRA learned. These are observations from their setup, rather than guaranteed results on other hardware.