Innovate Futures @ Benji

MiniMax H3 Character Swap LoRA — Replace Any Character in Video, No Masking (Patreon)

Published:

2026-10-03 13:00:22

Edited:

2026-10-03 16:06:50

Imported:

2026-10

Downloads

Download minimax_h3_character_swap_1.zip

Content

Video : https://youtu.be/koI_VOTe0nM

Hey everyone.

New video is up https://youtu.be/koI_VOTe0nM , and this one is for anyone who ever wanted to drop their own character into real video footage. It is about the new Character Swap LoRA for MiniMax H3. I will be honest — when I first saw it, I thought it was just another video-to-video swap effect. Then I tried it myself, and I got a bit addicted playing with it.

The whole thing is refreshingly simple. You feed it a character sheet image, a reference video, and basically one sentence telling it who to swap. No masking, no SAM, no blurry edges around the character. Because H3 is an Omni model, it just understands what you are pointing at.

In the video I test it on walking footage, a kick, a guy doing a backflip, and full action scenes. I even add fire and energy effects on top of a swap, and it holds together.

The one surprise for me: the outfit matters. Keep the character simple and the background stays untouched. But a complex costume can pull the whole theme of the scene along with it.

Update: For Patreon Supporters , your additional content here:

https://www.patreon.com/aifuturetech/posts/h3-singularity-171325006

Thanks for watching.

— Benji

Resources:

Video : https://youtu.be/koI_VOTe0nM

Calliope - Local-first Agentic Idea-to-video Studio https://github.com/benjiyaya/Calliope

- Character Swap LoRA: https://huggingface.co/akatz-ai/MiniMax-H3-Character-Swap-LoRA

- LoRA file: h3_character_swap_pro4500_1000.safetensors

- Latent upscaler: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler

Files

MiniMax H3 Swap Any Character in Video - No Masking Need

MiniMax H3 character swap, done with a single LoRA and no masking at all. This video walks through the new MiniMax H3 Character Swap LoRA in ComfyUI: the dedicated text prompt format, why complex outfits can leak into the background, and how simple character sheets keep the swap clean. If you build AI video in ComfyUI and want to replace a person in any footage with your own character, this covers the full pipeline. All you need is a character sheet image (I generate mine with Qwen 2.1), a reference video, and one sentence in the LoRA's prompt template. Because MiniMax H3 is an Omni model, there is no SAM 2 or SAM 3 masking step, no edge blur, no shadow residue. The ReferenceToVideo node takes the image and video directly. Character swapping used to mean heavy masking workflows and hit-and-miss inpainting. This LoRA changes that: res multistep sampling, beta scheduler for more texture, 8 steps, then a 1.2x to 1.5x latent upscale. I test it on walking footage, a kick, a backflip, and even energy effects added on top of the swap. One character per swap is the safe bet; two characters works sometimes but it is hit and miss. The video also explains why the reference mechanism needs the swap keyword spelled out in the prompt, not just the files plugged in. Calliope - Local-first Agentic Idea-to-video Studio https://github.com/benjiyaya/Calliope Resources: - Character Swap LoRA: https://huggingface.co/akatz-ai/MiniMax-H3-Character-Swap-LoRA - LoRA file: h3_character_swap_pro4500_1000.safetensors - Latent upscaler: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler Freebie Workflow: https://www.patreon.com/aifuturetech/posts/minimax-h3-swap-171306927 If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech #MiniMaxH3 #ComfyUI #AIVideo #CharacterSwap #LoRA