MiniMax H3 Person Remover V1 is an experimental adapter from Akatz Labs for removing a selected person from an existing video. Its October 8, 2026 announcement links to a public model and ComfyUI workflow. It is a community modification, not an official MiniMax release or an editing tool currently offered by H3Vid.
The useful question is how to prepare a shot and judge whether the removal worked. The following guide is based on the model card and workflow instructions. We have reviewed those sources but have not run the adapter ourselves. The cover is an AI-generated concept illustration, not a Person Remover result.
How removal differs from character replacement
Character replacement keeps a performer in the scene and changes their identity. Person removal asks the model to fill the area behind them. That makes background continuity the main acceptance criterion: a convincing empty corridor is useless if its floor lines drift or a second actor disappears.
The published workflow combines person tracking, a green mask, a clean first-frame reference, and H3 Ref2VA generation in overlapping windows. The LoRA does not provide segmentation or create the clean reference itself. Those are separate preparation stages.
This distinction also separates it from a character-swap workflow and from watermark removal. A tool that removes a logo does not necessarily know how to reconstruct a moving body’s background.
Prepare a small, inspectable test
Choose a continuous shot with one clearly identifiable person, a modest camera move, and a background you can recognize. Avoid starting with a montage: a cut makes it difficult to tell whether an error comes from tracking, reference preparation, or continuation.
Keep an untouched source and export a short working copy. Record its dimensions, frame rate, and frame count before editing. You will need that record when checking the final length and soundtrack.
The documented input uses 24 fps and dimensions divisible by 32. The release recommends beginning with roughly five seconds. Convert and crop deliberately rather than changing framing halfway through the workflow. If resizing is necessary, prepare the clean reference from that resized version so both inputs describe the same image coordinates.
Assemble the documented workflow
Download the adapter and the supplied Person-Remover-Window-Reroll-V1.json from the model repository. Put the adapter in ComfyUI/models/loras/. Use a current H3/SAM-capable ComfyUI installation and H3 Relay with the window preview and reroll controls.
The base model, text encoder, VAEs, and segmentation checkpoint are separate downloads. Follow the filenames and loader folders in the model card; possessing the approximately 155 MB LoRA alone does not establish that the rest of the runtime is installed. The author’s recorded RTX 4090 setup is a validation configuration, not a published minimum hardware requirement.
Before generating, inspect every loader for a missing model or incompatible attention backend. Save a copy of the workflow with your actual filenames. This gives you a useful recovery point if updating a node later changes its inputs.
Make the clean first frame convincing
Remove the selected person from the first frame using an image editor or an image-editing model. Keep the same size and framing. Pay particular attention to floor joints, wall edges, door frames, and objects partly hidden by the person.
A good still image should also fit the surrounding footage. Compare it with nearby frames: does a railing continue naturally, and does the reconstructed light match the original shot? An attractive invented background can still be the wrong reference for this video.
For an initial test, use a person standing in front of a simple wall or corridor. Treat that as a proposed test design, not a demonstrated capability. Save both the unedited first frame and the clean version so you can trace a later distortion back to its input.
Inspect the mask before spending time on generation
Describe the target person specifically, then run the mask preview first. A clothing-based description is easier to inspect than a vague instruction such as “remove someone.” Review the complete preview, including moments when the subject turns, crosses another person, or leaves the frame.
Look for missed hands, hair, reflections, and shadows. Also check that the selection has not jumped to another actor. Enlarging a mask may capture an edge, but it also asks the model to invent more of the scene; inspect that tradeoff before increasing it.
Generate, reroll, and check the rebuilt continuation
The release starts at a 22-frame window, 12 sampling steps, and LoRA strength 1.0. Begin with those documented settings rather than combining unrelated acceleration adapters immediately. Change one setting per trial and retain the seed and input record.
Window rerolls preserve an earlier prefix and rebuild the selected window and its continuation. Consequently, a better local repair can alter later footage. Watch the suffix after every reroll, especially the first few frames at each boundary. Compare camera direction, repeating textures, background objects, and remaining people with the source.
Do not judge only the best frame. Play the clip normally, scrub the edited region, and review transitions frame by frame. Reject a result if the removed person’s outline reappears, the floor swims, or an unselected person changes identity.
Audio, licensing, and delivery
The clean output is silent by default. The workflow guide explains how to reconnect original audio; this also preserves the removed person’s voice. Visual removal is not voice removal. Listen to the exported clip and verify synchronization after any trimming.
Review the weights’ license before local use. The model card states that its standard territorial grant excludes the US, EU, UK, and Republic of Korea and describes separate authorization. H3 Relay’s code license does not change the weights’ terms.
Finish with an export checklist: correct frame count, intended soundtrack, stable background, and no unintended changes outside the selected subject. The author’s showcase contains selected successes, so it cannot establish a general success rate. For broader H3 setup context, use the ComfyUI guide; for this adapter, keep the original model card and workflow guide beside your test record.