MiniMax H3 Character Swap: Workflow and Limitations

MiniMax H3 Character Swap LoRA is an experimental adapter from Akatz Labs for replacing a character in a source video using a reference image. It requires a compatible H3 base model and runtime. It is not a standalone generator or an official MiniMax character-swap service. Read the publisher’s model card.

Sources checked October 1, 2026. This is a documentation-based guide. Prompt examples are illustrative and have not been benchmarked by ZNIX; they are not generated-output samples.

Prepare the inputs and runtime

The publisher specifies an H3 Ref2VA-capable runtime, separate base weights and VAEs, and the released adapter. Its documented starting strength is 1.0. Use the linked workflow instructions for your version instead of assuming any H3 or H3 Max endpoint accepts a LoRA.

Use a source clip and character reference you have permission to edit. For a first trial, choose a single clearly visible character, a steady camera and a simple background. Avoid starting with a montage: it makes failures harder to diagnose.

A first replacement workflow

1. Keep an untouched copy of the source video. Note its duration and frame rate.
2. Prepare a clear replacement reference with the intended outfit and style. Remove unrelated subjects from the reference.
3. Load the compatible base workflow and adapter, following the publisher’s input mapping.
4. Supply the source as <Video 1> and the character reference as <Picture 1>. Identify the target unambiguously.
5. Run a short trial and compare it against the source before attempting a longer sequence.

Illustrative replacement prompt

In <Video 1>, replace the person wearing the green jacket with the illustrated robot in <Picture 1>. Keep the robot’s orange panels and rounded head. Leave the room, desk and other people unchanged. Follow the original person’s position and action.

This example expresses an editing goal; it is not evidence that the adapter will preserve every detail. If the target is ambiguous, simplify the clip or describe the person more precisely rather than adding unrelated style instructions.

What the publisher reports

The model card describes more promising results on roughly four-to-five-second continuous shots than on longer tests. It reports unreliable timing, expressions and hard cuts. Copying source audio back onto an edited video is postprocessing, not proof of generated audio fidelity. These are publisher observations, not a ZNIX benchmark. Read the findings and limitations.

Review the result in three passes

Identity: compare the replacement at the beginning, middle and end. Check costume details, head shape and whether the original person reappears.

Scene preservation: compare stationary objects and people who were not selected for replacement. A convincing character is insufficient if the room or another person changes.

Timing: watch the source and result side by side. Look for shifted gestures, missing actions and altered cuts. Listen with the final audio track attached, especially when speech is visible.

Record each failed criterion before changing settings. If the source itself is complex, reduce the shot to one action; more instructions can make the request harder to satisfy.

Check compatibility before choosing a hosted tool

The repository points to the upstream MiniMax H3 Community License. Follow those terms and the exact runtime requirements before using the adapter in a production workflow. Availability of a hosted MiniMax model alone does not confirm adapter support.

For context on a different model offering, see our H3 Max overview; its hosted options should not be interpreted as support for this LoRA. If your task is to generate a new scene instead of replacing a character, read the LTX 2.5 guide or explore ZNIX’s video tools.