Introduction

Ever nailed a shot, only to have a stranger wander right through the middle of it? Akatz Labs just released MiniMax H3 Person Remover LoRA V1, an open-weight video-to-video adapter that removes a chosen person from a clip and rebuilds the background behind them. It runs locally in ComfyUI on top of MiniMax H3 Ref2VA, and it ships with a ready-made workflow, the training dataset, and the full training recipe.

The repo went live on Hugging Face on October 8, and the launch post quickly passed 600 likes and roughly 700 bookmarks. Clearly a lot of creators have a photobomber or two they'd like to evict.

What shipped

  • Model: H3 Person Remover V1, a LoRA for MiniMax H3 Ref2VA (pruned)
  • Maker: Akatz Labs (akatz-ai), the team behind the popular H3 Character Swap LoRA
  • Modality: video-to-video person removal with background reconstruction
  • Weights: open download, 155 MB safetensors file, rank 16
  • Where to run it: your own ComfyUI with native MiniMax H3 and SAM 3.1 support, plus the H3 Relay custom nodes
  • Extras: the Person-Remover-Window-Reroll-V1 workflow JSON, the H3-Person-Remover-v1 dataset, and the training configs

How it works

The trick is a tidy little pipeline rather than one magic button:

  1. SAM 3.1 finds the person. You type a short description such as "man in gray shirt", and SAM tracks that person through the whole clip.
  2. The mask goes green. H3 Relay expands the mask by five pixels and fills it with pure green, so the model knows exactly what to replace.
  3. You supply a clean first frame. Remove the person from frame one yourself, in any image editor or image-editing model, keeping the same size and framing. This frame is the model's anchor for what the empty scene should look like.
  4. H3 rebuilds the background in windows. The workflow generates the clip in overlapping chunks (22 frames by default). Each new chunk uses the last generated frame as its reference and carries 18 frames of history, then H3 Relay stitches everything back to the original length.

The nicest touch is the window reroll. Every finished chunk appears as a preview card with its own seed, a Lock button, and a Reroll button. If one section looks off, you can redo just that window and everything after it, without regenerating the whole clip.

Starting settings

The model card recommends these values for a first run:

  • LoRA strength: 1.0
  • Window length: 22 frames (H3 lengths follow 17n + 5, so 22, 39, 56 and so on)
  • Sampling steps: 12
  • Sampler and scheduler: er_sde / simple
  • CFG: 1
  • Video and audio sigma shift: 12 / 3
  • Mask expansion: 5 pixels
  • Input: 24 fps, width and height divisible by 32, ideally a continuous shot of about five seconds

No Turbo or VFX LoRA is needed. Longer windows use more memory and aren't necessarily better, so start with 22.

Quick install

Drop the LoRA into your loras folder:

cd ComfyUI/models/loras
wget -O H3-Person-Remover-V1.safetensors "https://huggingface.co/akatz-ai/MiniMax-H3-Person-Remover-LoRA/resolve/main/H3-Person-Remover-V1.safetensors?download=true"

Then install or update H3 Relay to a version that includes Person Remover reroll support, drag the example workflow into ComfyUI, and select the base files listed on the model card: the H3 Ref2VA pruned INT8 ConvRot diffusion model, the Qwen3VL 32B NVFP4 text encoder, the H3 video and audio VAEs, and SAM 3.1 multiplex FP16.

The default removal prompt, ready to paste:

Remove the green-masked person and reconstruct the background. Preserve the rest of the video, including its camera motion and frame timing.

Tips and limits

  • Keep the original sound. The clean output is silent by default. Connect Get Video Components audio to Create Video audio to keep the soundtrack. Note that this also keeps the removed person's voice.
  • Check the mask first. Preview the green mask before a full run, and watch hands, hair, shadows and reflections. Anything the mask misses can stay behind.
  • Your clean frame matters. A sloppy first frame carries through every later window.
  • The whole scene is regenerated. Areas outside the mask aren't pixel-locked, so expect small drift. Strong camera moves and hard cuts can also break continuity.
  • It's experimental. Akatz says the showcase is a set of selected successes, not a benchmark. The validated setup was an RTX 4090 with ComfyUI 0.37.0, which is a tested configuration rather than a minimum spec.

License check

The LoRA is distributed under the MiniMax H3 Community License Agreement, the same terms as the base model. Its standard territorial grant excludes the US, EU, UK and Republic of Korea, and it has its own use, redistribution and commercial rules, so read the LICENSE before using it on client work. H3 Relay is a separate GPL-3.0 project.

What's next

Akatz is already using Person Remover to build cleaner training samples for Character Swap V2, and says an updated Character Swap workflow for longer videos is coming next. Removing someone is the first half of replacing them, so this pair could make H3 a seriously handy local VFX kit.

Original Source

Model card and downloads: MiniMax H3 Person Remover LoRA on Hugging Face

Custom nodes: H3 Relay on GitHub

Training data: H3-Person-Remover-v1 dataset

Conclusion

Person Remover turns one of the oldest editing chores into a short ComfyUI run: describe the person, hand H3 a clean first frame, and reroll any window that wobbles. It isn't flawless yet, but for short, steady shots it's a delightful new tool for anyone running H3 at home.

Want to play with MiniMax H3 without a local setup? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3-person-remover-lora-comfyui-video-cleanup

—Aurelia ♡