MiniMax H3 Video-to-Video — Character Swap, Clothing Swap & Video Editing Without Masking [video post] (Patreon)
Videos
-
vid_00050.mp4 -
vid_00004.mp4
Downloads
Content
Video : https://youtu.be/3LdtyMeELWk
Hey everyone.
New video is up, and this one is about video-to-video with MiniMax H3. Not the old ControlNet-only approach — this is prompt-driven V2V where text prompts do the heavy lifting and ControlNet is optional.
To Patreon supporters, some people might no know, and usually 1 video post for everyone like this , and sometimes if something more advance I want to go more in-depth, I will create Patreon post that is related to it, and it will have a related links to the video post.
So, read, before make your own wrong assumption, that make lives better. This message is not to all people, so don't put yourself in if you are not in this case.
Continue.
The big difference from how we used to do video-to-video: instead of fully relying on ControlNet to drive everything, H3 wants you to work with long-form text prompts. You define Subject 1 (your character), Subject 2 (your background environment), and Video 1 (your reference video), then let the model figure out how to blend it all together. ControlNet is there if you want it — Depth Anything V2, DWProcessor, or both — but in some cases I bypassed ControlNet entirely and just used text prompts.
I tested seven scenarios: TikTok dance character swap with transparent background images, clothing swap where arms cross over the new outfit naturally, concert stage video editing with zero masking or segmentation, drone view effects with camera push-in via text only, podcast-style single character swap, and dual-character swapping. The concert stage one surprised me the most — the fence separating the fire from the stage came out so clean it does not look like any masking technique I have used before.
I also demo the ComfyUI-H3-VisionPromptor custom nodes I just released. Vision Prompter generates H3-formatted text prompts from your input images, and Vision Analyzer does in-depth image captioning. I show how to use both with Qwen 3 VL to build structured prompts without writing essays by hand.
Links from the video:
MiniMax H3 Turbo (lightx2v) — https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
MiniMax H3 ComfyUI LoRAs (Kijai) — https://huggingface.co/Kijai/MiniMax-H3_comfy/tree/main/loras
ComfyUI-H3-VisionPromptor — https://github.com/benjiyaya/ComfyUI-H3-VisionPromptor
— Benji
Update: MiniMax H3 V2V Multi-Pass Pipeline — For Patreon Supporters
Freebie workflow attached: