Home Creators Posts Import Register Favorites Logout

Videos

  • vid_00010.mp4

Downloads

Content

Hey everyone.

New video is up. The Character Swap LoRA for MiniMax H3 has been updated, and this one is for anyone who tried the last version and lost the face. Same file, same filename — the author just posted the update with a new hash, and I tested it end to end.

The question the last version left you with was: can it swap a character and keep the expressions? Now yes. The swap follows the reference clip's facial expressions, and eye contact holds up even when the person looks off-camera at different angles. I also threw harder shots at it — clips where objects cross the face, a selfie-style shot with makeup over the skin — and the objects survive the swap, which is what sells it as real.

The part that surprised me: it does all of this with no masking. The swap edges stayed clean on every clip I ran, including close-ups.

The free single-pass workflow from the video is attached below, grab it and run it. Supporters, the two-stage version with the latent upscale and refine pass is in the supporter post.

Patreon Supporters check out the member post for additional contents for you. https://www.patreon.com/aifuturetech/posts/172382699

— Benji

Resources:

LoRA: https://huggingface.co/akatz-ai/MiniMax-H3-Character-Swap-LoRA/blob/main/h3_character_swap_pro4500_1000.safetensors

Files

Minimax H3 Local AI Video Character Swap LoRA With Facial Expressions

Character Swap LoRA for MiniMax H3 is updated to follow facial expressions, and this ComfyUI tutorial tests the new version end to end. Using the akatz h3_character_swap_pro4500_1000 LoRA with the Singularity ref2VA model and Veda sparse attention for a small speed boost, we swap characters video-to-video in ComfyUI while keeping pose, scale, and now the facial expressions of the reference clip. This one is for ComfyUI users and AI video creators who tried the previous character swap update and wanted it to hold up on faces. The prompt format is the same one from the earlier tutorial, so if you already have the workflow set up, you just pull the updated LoRA file (same filename, new hash) and run it. No masking required — the swap edges stay clean even on close-up shots. Why it matters: the last version swapped bodies well but lost expressions. This update follows eye contact across camera angles, keeps objects overlapping the face intact, and produces clean enough output that a single sampler at 0.5-0.7 megapixels already looks usable. Tested on 8-10 second clips of talking characters, including a selfie-style shot with makeup and foreground objects. Resources: - Character Swap LoRA (updated, facial expressions): https://huggingface.co/akatz-ai/MiniMax-H3-Character-Swap-LoRA/blob/main/h3_character_swap_pro4500_1000.safetensors Workflow for Freebies: https://www.patreon.com/aifuturetech/posts/172382070?utm_source=youtube&utm_medium=video_description&utm_campaign=h3_character_swap_facial_expressions&utm_content=patreon_public_post Calliope - Local-first Agentic Idea-to-video Studio https://github.com/benjiyaya/Calliope If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech #ComfyUI #MiniMaxH3 #aivideo

Comments

Med Boussouni

Great Workflow ! Many Thanks ! do you plan a multi character workflow ?

Innovate Futures @ Benji

the way Multi, is not suitable for this LoRA. That is a loop of generate handle each character one at a time.

Bikram Saha Ray

Your workflow works but beyond 5 second everything goes haywire.. I tried to generate in chunks and stitch, but beyond 5 sec nothing holds.. how can I get it to work for 10 - 15 seconds clip

Innovate Futures @ Benji

see my demo I am doing 10 -15 seconds. as well as other people in our Discord can do it. So your question have to be specific what is the problems identified it and ask that way can be solve, otherwise this is a very broad question.

Bikram Saha Ray

For the first 5 seconds the character swap works perfectly.. but after that the video becomes weird, either the person in the video does something completely different or the character from ref image does some weird moves.. I am running it on rtx 5070 ti 12gb vram

Innovate Futures @ Benji

You've actually diagnosed it correctly yourself: "the first 5 seconds work perfectly, then it goes weird." The 5-second point isn't your prompt or the LoRA — 124 frames (~5.17 s) is the default window length, and on a 12 GB card a 10–15 s pass in one go won't fit in VRAM. Once it doesn't fit, ComfyUI falls back to lowvram/offload, and with a GGUF checkpoint the GGUF LoRA patching can quietly stop applying — which is exactly "the swap fades and the character drifts back to the original." I can do 15 s in one pass because I'm running a 96 GB workstation card — that doesn't transfer to 12 GB, and I should have said that at the start instead of asking you to be more specific. Three things, in order: 1. Move off h3_character_swap_pro4500_1000 to h3_character_swap_v1.5 at strength 1.0. It's trained on moving targets, not just stills, so it holds identity well past the first few seconds. That one change should fix most of the "weird moves." 2. Don't stitch independent chunks. Each chunk re-rolls identity from nothing. Carry the last ~18 frames of the previous window in as history (akatz's H3 Relay "Windowed Edit" node does this automatically), and re-declare in every chunk prompt — references aren't inherited between segments, so if chunk 2 doesn't name the image, it simply doesn't get it. 3. Drop to ~864×480 and raise steps to 20–25 instead of 4-step turbo when identity has to survive motion. Turbo is fine for testing, but it weakens reference adherence. Also: update ComfyUI-GGUF — older builds had a real bug where lowvram made LoRA loading "almost have no effect." That alone can look like your problem. Post your workflow if you want and I'll check the frame settings — worth confirming your length is landing on the 17n+5 grid too.