Skip to main content
Back

fal Fixes MiniMax H3's Plastic Faces With an Open Realism LoRA

fal launched a MiniMax H3 LoRA trainer and shipped Realism People, an open-weight adapter that sharpens skin, faces, and human motion.

Model, Full Access
  • fal released Realism People, an open LoRA for MiniMax H3 focused on photorealistic humans.
  • Trigger word r34l1sm, rank 32, trained on 176 hand-curated clips at 24 fps.
  • Works with text-to-video, image-to-video and reference-to-video pipelines out of the box.
  • fal also opened its H3 LoRA trainer across four task-specific endpoints.
  • Trainer pricing starts at $0.005 per step for text-to-video, $0.01 for other modes.
  • Training resolution mattered more than rank or step count for fine skin detail.

Video models keep getting sharper on landscapes and product shots, but people are where they break. Faces plasticize in close-up, skin loses texture, micro-expressions drift. fal just released a LoRA that targets exactly that failure mode on MiniMax H3, and opened up the trainer used to build it, so anyone can now fine-tune H3 on their own footage.

MiniMax H3 itself is worth a quick refresher. It is a 33B dense transformer released under the MiniMax H3 Community License. It reads text, images, video and audio as a single context and returns video with native stereo sound at up to 2K resolution and 15 seconds in length. Text-to-video, first-and-last-frame, reference-to-video, and precise video editing all route through one endpoint rather than a stack of task-specific experts.

What the Realism People LoRA actually changes

The adapter is deliberately narrow: a LoRA for MiniMax H3 specialized in realistic humans. Faces that hold up in close-up, natural skin texture, believable expressions and gestures, film-style lighting, documentary camera movement. Same prompt, same seed, same trigger word on both sides, base on the left and adapter on the right, is the comparison fal published, so the adapter is the only variable.

The Realism People LoRA succeeds an earlier Realism release, retrained on a larger people-focused corpus. It keeps H3's native synchronized audio and works across the same three modes as the base model. Skin holds pores instead of smoothing, eyes stay coherent, and motion picks up a subtle handheld quality that plain H3 tends to iron out.