fal MiniMax H3 Realism People LoRA Brings Photoreal People to ComfyUI

fal released the MiniMax H3 Realism People LoRA on August 10, an open-weights adapter specialized in photorealistic people. It is the successor of fal's earlier MiniMax-H3-Realism-LoRA, retrained on a larger dataset focused on people, and ships as a single 125 MB safetensors file that covers text-to-video, image-to-video and reference-to-video generation with the trigger word r34l1sm.

One prompt, one seed, sixteen LoRA configurations trained while building this adapter (image from the official model card)

What it does

MiniMax H3 is already a strong general video model. This adapter pushes it further on human-centered shots: faces that hold up in close-up, natural skin texture instead of smoothing, coherent eyes and micro-expressions, film-style lighting, and a subtle handheld camera quality, while keeping H3's native synchronized audio.

The model card backs the claims with 19 before/after pairs using the same prompt and the same seed — the only variable is the LoRA. The comparison video below plays each pair with the base model first, then the adapted version beside it:

197-second before/after comparison from the official model card: close-up talking faces, arguments, weathered skin, children, ritual and travel scenes

How to use it in ComfyUI

The weights are a standard safetensors LoRA, so no custom nodes are needed:

1. Download h3-realism-people-t2v-i2v-r2v.safetensors from the model repository into ComfyUI/models/loras/.
2. Add a LoRA loader to your MiniMax H3 workflow (T2V, I2V or Ref2V) and point it at the file.
3. Start the prompt with the trigger word r34l1sm and describe the scene. A scale of 1.0 is the intended strength; lower it to 0.6–0.8 for a lighter touch.

Availability

The LoRA is free and open-source on Hugging Face (fal/MiniMax-H3-Realism-People-LoRA, 125 MB, T2V/I2V/R2V in one file). It is also served through fal's LoRA endpoint at fal.ai/models/minimax/h3/text-to-video/lora for API users.