Home Creators Posts Import Register Favorites Logout

Videos

  • 764400898.mp4

Downloads

Content

Hey everyone.

If you watched the Character Swap update video https://www.patreon.com/aifuturetech/posts/character-swap-172382070 , you saw the new expressions, the eye contact test, and the object-over-face shots. This post is the run-manual for the two-stage version that produced those results — download the files below, and you can reproduce every clip from the video without rewatching it.

----------------------------------------

WHAT TO DOWNLOAD (MODEL FILES)

----------------------------------------

Base model (ComfyUI/models/diffusion_models/minimax-h3/):

Minimax-h3_Singularity_ref2va_v1.3_int8.safetensors

Text encoder (models/text_encoders/minimax-h3/):

qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors

VAEs (models/vae/minimax-h3/):

minimax_h3_video_vae_fp16.safetensors

minimax_h3_audio_vae_fp32.safetensors

LoRAs (models/loras/minimax-h3/):

h3_character_swap_pro4500_1000.safetensors — the updated character swap, strength 1

minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors — turbo, strength 1

h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors — realism, strength 1

Veda predictor (models/veda/):

minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors — same one from the Veda video

Upscale stage (models/upscale_models/ or models/diffusion_models/minimax-h3/):

minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors

Frame interpolation (models/frame_interpolation/):

film_net_fp16.safetensors

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

Everything else: https://huggingface.co/Comfy-Org/MiniMax-H3

----------------------------------------

QUICK START

----------------------------------------

1. Load MinimaxH3_CharSwap-2Stage-Advance-20261015.json.

2. Check every loader points at your own folder names — the paths in the file follow mine.

3. Load your reference video (the clip you are swapping inside) and your character image (the replacement).

4. Type the prompt using the format below.

5. Set the frame count to your clip length on the 17n+5 grid — 192 frames is 8 seconds at 24fps.

6. Seed ships fixed, resolution 1344x768. Queue it.

That's the whole loop.

----------------------------------------

WHAT TO INPUT (PROMPT FORMAT)

----------------------------------------

Same format as the last character swap video — plain English, no trigger word. You name the target person in the video by description, point at Picture 1, and tell it what to preserve:

Replace only the woman with brown hair in <Video 1> with the character in <Picture 1>.

Keep the face, identity, outfit and art style from <Picture 1>.

Preserve the camera, background, lighting, and the character's pose, scale,

facial expressions, eye contact and movement.

Rules: describe the target by what you can see (hair, clothing, position in frame) — if two people are in the shot, be specific or you will get both. The picture goes in through the reference image input, the video through the reference video input, so the <Video 1> and <Picture 1> labels just need to match those inputs. Do not add "show the reference image" wording, the swap works from the native reference path.

----------------------------------------

KEY SETTINGS (AND WHY)

----------------------------------------

Stage 1 (base pass): euler, beta scheduler, 8 steps, denoise 1.0. This is the blurry first pass you see in the video — that is expected, it is just producing the latent for stage 2.

Stage 2 (refine): sa_solver, simple scheduler, 1 step, denoise 0.2, seed 777. One low-denoise step on the upscaled latent — enough to sharpen without touching the swap or the expressions. This is the whole reason the two-stage version exists.

Latent upscale chain: ships at 1280x704 fp16 with the ultimate upscale and spatial split values I tuned. Leave the whole chain as shipped — the split tile settings are the difference between this and the free workflow's flat output.

Character swap LoRA: 1.0. Turbo 4-step: 1.0. Realism: 1.0. If the realism look is too strong for your clip, drop the realism LoRA to 0.7 — that is what last month's workflow shipped at.

Veda: t2va 8nfe predictor, tile size 32/32, leave as shipped. Sigma shift 12/3 and chunk feed forward 2/4096 also ship correct, do not touch them.

Resolution and duration: my test renders were 0.5 megapixels at 8 seconds for the raw single-sampler check, and 0.7 megapixels at 10 seconds for the eye contact test. Test low, render at 1344x768.

Frame interpolation (2x) and DLSS enhance are on by default. Turn DLSS off if you want a softer, more film-like output.

----------------------------------------

TIPS

----------------------------------------

1. Re-download the LoRA even though the filename never changed. The update rides on the hash — if your copy still behaves like the old one, your copy is old.

2. Do not mask. The edges held on every clip in the video, close-ups included. Masking was the workaround for the old version, it is wasted time now.

3. Run the free single-pass workflow first as a test loop. If the swap and expressions look right there, the two-stage render will only add sharpness.

4. Objects crossing the face survive the swap — hands, cups, makeup, hair. Clips with foreground objects are the ones that make the result look real instead of pasted.

5. Eye contact is the real test of this update. Grab a clip where the person looks at someone off-camera from an angle and check whether the gaze holds. It does now. It did not before.

6. The first pass looking blurry is not a bug. Stage 2 exists because stage 1 is only setting up the latent.

----------------------------------------

COMMON ISSUES

----------------------------------------

Expressions not following the reference clip — you are running the old LoRA file. Re-download it, the hash is the update.

First pass looks blurry — expected, that is the stage 1 latent. Let stage 2 finish, or check the refine seed is still 777 and denoise 0.2.

Swap edges smearing on fast motion — drop the realism LoRA to 0.7, and if it persists, raise stage 1 steps from 8 to 12.

Face identity drifting on side profiles — make the target person description more specific: hair color, outfit, where they stand in frame.

VRAM overflow at higher resolutions — test at 0.5 megapixels first, keep chunk feed forward at 2 and Veda on, both are already shipped that way.

Output looks over-sharpened or plastic — turn the DLSS enhance node off and re-run stage 2 only.

----------------------------------------

Everything is attached. Grab the LoRA fresh from the link above before your first run.

Questions? Drop them in your comment with your GPU, VRAM, and the exact error.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji

Files

Comments

Makari Oh

Do all the models need to be downloaded if we already have the first version (members version)? Or do we need just the new updated swap lora? Thanks for this 🙏

Makari Oh

Im confused. My agent is saying the lora is exactly the same as the pre-updated version. Same weights/hashtags?

Makari Oh

Ive had no issues with facial expressions with the pre-updated workflow. Smiles, eye movements, open and closed mouth...etc. all work with the first release for me Lol

Innovate Futures @ Benji

oh nice, because my last post try this Lora , that half robot man, it can't make any lipsync, but this time I can. LOL. And saw the author said, it updated.

Makari Oh

To be honest I've tested both releases and I get better results from the first release 🤔 but I guess it depends on the machine running the workflow? Anyway I love this character swap workflow. It works really well 👍

Makari Oh

Can I use uncensored text encoder for updated workflow? Or has to be the one you listed in new post?

Garrulous Skeptic

In the 2ns Sampling + refiner I'm getting a red highlight - do i need to change something - i bypassed to get it working. Same with the DLSS5Settings. I have a 4090. Thanks

Renan Bastos

consigo usar o fluxo no Runpod? Linox?

arquinovatos

Hello, thank you so much for your hard work! I can't find the DLSS5 nodes that you used I noticed I can't generate videos longer than 8 seconds as it does not swap the character anymore, are you willing to bring loop workflows as with wan?