MiniMax H3 Character Swap LoRA Recasts Anyone in a Reference Video

Source video frame	Character sheet reference	Dataset target: the character swapped in

A training example from the published dataset, shown as the source frame, the character sheet and the edited target the LoRA was taught to produce. Source: H3 Character Swap v1.

The task is wider than identity transfer. The H3 FaceSwap LoRA moves a face onto the performance already in the clip; this adapter copies the whole replacement character, including its outfit and its rendering medium, while preserving the source scene's medium in turn. That is what makes cross-style swaps possible: an illustrated character dropped into live-action footage keeps its illustrated look, and 34 of the 94 training edits were built exactly that way.

What the dataset contains

Akatz Labs trained the adapter on H3 Character Swap v1, a reference-conditioned dataset published alongside the weights:

• 134 examples: 94 character-swap edits and 40 preservation clips. Training used 76 edits and 32 clips, with 18 edits and 8 clips held out.
• Edits are built as one still image each. The scene ships as a five-frame static video (<Video 1>), the replacement arrives as a character image or sheet (<Picture 1>), and each pair carries a short targeting instruction such as Swap the man in the purple shirt in <Video 1> with the character in <Picture 1>.
• Cross-style and sheet coverage. 34 edits swap between rendering styles and 20 use complete character sheets, with portrait, square and landscape sheets kept at their original aspect ratio.
• Regularization clips are explicit preservation examples. The same video is used as both control and target with the caption Keep <Video 1> unchanged.
• One character at a time. Multi-character replacements were not supervised in the training targets.

The adapter itself is a rank 16 LoRA, 148 MB, trained for 1,000 updates on the pruned INT8 Ref2VA base from Comfy-Org/MiniMax-H3 with the Ostris training assistant. Only the final checkpoint is published, and the training launcher and run configuration ship in the repository under training/.

More examples

Source video frame	Portrait character reference	Dataset target: the character swapped in

A second training pair, this one conditioned on a portrait reference rather than a sheet.

What it does well, and where it breaks

The author's own comparison against the base model is narrow and worth reading before downloading:

• Scene preservation improved. In side-by-side reviews the adapter kept backgrounds and scene structure closer to the source than the base model, but the author labels those reviews qualitative rather than a benchmark.
• Long clips drift. Framing, placement and source timing wander over longer windows, and hard cuts can degrade into slow repositioning instead of an edit.
• Expressions are unreliable. Close-up facial expressions often do not follow the original performance, and stronger expression instructions sometimes suppressed the swap entirely.
• Short clips are the sweet spot. Continuous shots of roughly 4 to 5 seconds held up better than the author's full 14-second attempts; no precise maximum duration has been established.
• No trigger word. Targeting comes from the prompt, and the card is explicit that prompt wording is not a guarantee of strict source alignment.

Banodoco testers hit the same wall. Charlie ran a ten-second test with the adapter on a hybrid FL2VA/Ref2VA H3 model and reported that it held to about five seconds, with a 15-second attempt turning into a blurred frame, while another tester noted that the pose and action of the original person still have to be described in the prompt or the swap does not copy the movement.

The reason for the drift is structural: because every edit target was a still image, the model never saw a moving character-swap target showing how to hold a performance across time. Akatz Labs applied to Banodoco's compute grant programme with that diagnosis and was approved for 40 H100 hours to shoot 12 to 24 moving video pairs covering expressions, cuts, occlusions and longer sequences, plus a comparison of the current reference conditioning against explicitly aligned source and target latent positions.

Running it in ComfyUI

The adapter is a single model-only LoRA with no node dependencies:

1. Download h3_character_swap_pro4500_1000.safetensors into ComfyUI/models/loras/.
2. Build or open a MiniMax H3 Ref2VA workflow and point it at the pruned INT8 Ref2VA base, its text encoder and the H3 video and audio VAEs.
3. Apply the LoRA at strength 1.0 through a model-only LoRA loader.
4. Load the clip you want to keep as <Video 1> and the replacement as <Picture 1>, then name the target person in the prompt, for example:

Replace only the man in the purple shirt in <Video 1> with the character in <Picture 1>.
Keep the replacement character's identity, outfit, and art style from <Picture 1>.
Preserve the source video's camera, background, lighting, objects, and all other people.
Match the target person's position, scale, pose, and movement.
Do not show the reference sheet or its background.

The card recommends 24 fps with H3's supported frame grid and short continuous shots over long ones. It notes the adapter does not require a Turbo LoRA, Spectrum or Sol attention to work, and that combining it with an 8-step Turbo LoRA was an evaluation choice rather than a compatibility claim.

Availability

• LoRA weights: akatz-ai/MiniMax-H3-Character-Swap-LoRA

• Training dataset and provenance: akatz-ai/H3-Character-Swap-v1
• Training configuration and launcher: training/ in the weights repository
• Base model: Comfy-Org/MiniMax-H3