☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1

An invented character, the same face again in any scene, and video of them, on one 12 GB card. Four workflows in one download: two new ones on MiniMax H3 Image (core ComfyUI nodes only) and the two video workflows I used with them. Every prompt from the video is in the zip.

Detailed step-by-step guide on GitHub: https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/ai-influencer-12gb

The four workflows

• 1 · AI Influencer 1, create the character: a text prompt gives you the first picture of an invented person.

• 2 · AI Influencer 2, same face, new scene: their close-up plus a prompt puts the same person somewhere else. With a second picture it puts them into a frame of your own video (the keyframe for the swap).

• 3 · AI Influencer 3, Viggle Remake swap: a clip plus one painted keyframe gives the same clip with your character, the clip's sound kept. The graph I ran, built on Viggle Remake.

• 4 · MiniMax H3 Ultimate 12GB: a still to a 10 second clip with sound, and the last frame of a clip as the first frame of the next for longer ones.

Workflows 1 and 2 are enough for an account that posts pictures. 3 and 4 are optional, only for video.

What you need

✔ Custom nodes: none for workflows 1 and 2 (core ComfyUI only). Workflow 3 needs Viggle's node pack (folder comfyui/ComfyUI-Viggle-Remake in Viggle/Viggle-Remake).

✔ ComfyUI version: 0.45 or newer. Older versions do not have the MiniMax H3 Image node. 0.45.0 tested.

✔ VRAM: 12 GB, tested, not estimated. Peak 11.6 GB.

✔ System RAM: 16 to 22 GB peak, in a 32 GB machine.

✔ Tested on: RTX 3060 12 GB · Ryzen 5 5600X · 32 GB DDR4 · ComfyUI 0.45.0

✔ Licence: workflows MIT · MiniMax H3 Image and MiniMax H3: MiniMax community licenses (check them before commercial use).

Models to download

Click a link to download the file. Put it in the folder shown, inside ComfyUI/models/.

Required for workflows 1 and 2: 3 files, 29.4 GB

File: minimax_h3_image_pruned_w4a8.safetensors (12.5 GB)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/diffusion_models/minimax_h3_image_pruned_w4a8.safetensors
Put it here: ComfyUI/models/diffusion_models/

File: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7 GB, the same file MiniMax H3 video uses)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Put it here: ComfyUI/models/text_encoders/

File: minimax_h3_image_vae_fp16.safetensors (1.2 GB)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/vae/minimax_h3_image_vae_fp16.safetensors
Put it here: ComfyUI/models/vae/

Optional, only for video (workflows 3 and 4)

File: viggle_remake_swap_pruned_int8_convrot.safetensors (21.0 GB, workflow 3)
Download: https://huggingface.co/Viggle/Viggle-Remake/resolve/main/int8/viggle_remake_swap_pruned_int8_convrot.safetensors
Put it here: ComfyUI/models/diffusion_models/

File: minimax_h3_video_vae_fp16.safetensors (5.2 GB, workflow 3)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
Put it here: ComfyUI/models/vae/

File: minimax_h3_audio_vae_fp32.safetensors (0.6 GB, workflows 3 and 4)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
Put it here: ComfyUI/models/vae/

File: minimax_h3_fl2va_pruned_int8_convrot.safetensors (21.0 GB, workflow 4, image to video)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
Put it here: ComfyUI/models/diffusion_models/

File: minimax_h3_video_vae_int8_convrot.safetensors (2.8 GB, workflow 4)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors
Put it here: ComfyUI/models/vae/

File: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors (2.0 GB, workflow 4)
Download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
Put it here: ComfyUI/models/loras/

Workflow 4 has more modes than image to video. The files for those are listed in downloads.md in the zip and on GitHub.

Step by step

1. Create the character (workflow 1): describe the person like a photo caption: who, hair, glasses, jewellery, outfit, place, light, lens. Long prompts work best. 3:4 at 7.5 MP is what I used. The same prompt with another seed is another person.

2. Make the reference (workflow 1 again): a close-up with both eyes visible. A half-body shot gives the face too few pixels, and every scene made from it looked worse in my tests.

3. The same face in a new scene (workflow 2): the close-up as Picture 1, Picture 2 stays bypassed. Start the prompt with "The woman from <Picture 1>, the same person with the same face..." and then describe the new place and outfit.

4. A reference sheet (workflow 2): the same, with a prompt that asks for four views on one picture: front, side, behind and a close-up.

5. Video from a still (workflow 4, optional): image to video with a scene picture as the first frame. I ran 576x736, 10 seconds, turbo LoRA.

6. Swap the character into a clip (workflows 2 and 3, optional): cut the clip to 90 to 345 frames at 24 fps on the grid 17n+5. Take one frame, paint the character into it in workflow 2 (Picture 1 = the frame, Picture 2 = the close-up, Ctrl+B switches Picture 2 on), then run workflow 3 with the clip and that keyframe. The clip's own sound is kept.

Speed on the RTX 3060 12 GB (cold start, model loading included)

Workflow 1, text to image: 3 MP 0:36 · 7.5 MP 0:51 to 0:54 · 16 MP 1:54

Workflow 2, one picture, 7.5 MP: 1:46 to 1:48 · two pictures (keyframe, 2880x1632): 2:13 to 2:22

Workflow 3, Viggle Remake swap: 832x480, 124 frames 6:37 · 1056x608, 90 frames 8:34

Workflow 4, image to video, 576x736, 10 seconds: 7:55 to 9:05 · three clips in a row (30 seconds): 24:19

Render the swap bigger than the 480 benchmark: faces are small at that size, and 608 looked clearly better on the same card. The quality is your call.

What did not work

At 16 MP something went wrong in the generation on my setup (same prompt and seed as the 7.5 MP picture), so I stayed at 7.5 MP. At 3 MP the face gets soft.

A half-body shot as the reference gave worse scenes than a close-up.

Small things drift between pictures. In mine: the shape of the sunglasses, and tattoos on the hands in one scene.

Over three clips in a row (30 seconds) the face drifted at the end.

Out of memory: lower the size, or for video make the clip shorter.

Where the line is

Good uses: UGC-style content, a product video, an ad for your own things, a character for your channel. Label it as AI when you post it. If you swap a character into somebody else's clip, say so. The character in the samples is invented; no real person's photo was used as a reference.

Get the workflow

GitHub (guide, prompts, all four files): https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/ai-influencer-12gb

More videos on this stuff: https://www.youtube.com/playlist?list=PLG_da2oqTx7k