Load one photo of a person, and MiniMax H3 draws a character sheet: the same person from the front, the three-quarter view, the side and the back, side by side in one picture. If your photo leaves something out, like the legs in a half-body photo, a box lets you name it, and the views show it. Switches make it faster, make it twice as big with SeedVR2, save each view as its own picture, or put your photo beside the sheet.
How to use it
Load your picture: one person, clearly in view.
Optional: type a background, like "a sunny beach". Left empty, it is plain light grey.
Optional: name what your photo doesn't show, like "blue jeans and white sneakers" for a half-body photo.
Run. The first run takes a few minutes, because about 42 GB of models load once.
What you get
One picture, 3072 × 1536: four views in a row, each 768 × 1536.
The views are framed like your photo. A half-body photo gives half-body views, a full-body photo full-body views.
To see more than your photo shows, name it in What the photo doesn't show. "Blue jeans and white sneakers" turned a chest-up photo into head-to-toe views in that outfit. Words like "full body" alone changed nothing.
The back is H3's guess from what your photo shows.
The switches
Fast: 10 steps instead of 18. The sheets looked just as good in my tests.
Upscale 2x: SeedVR2 makes the sheet twice as big on each side, 6144 × 3072, and sharpens it. About 25 seconds more.
Save each view too: the four views are also saved as pictures of their own, 768 × 1536 each (1536 × 3072 with Upscale 2x).
Your photo on the left: your photo goes on the left of the sheet, at the same height, with a thin white line between. The views from Save each view too stay without it.
What you need
ComfyUI 0.30.0 or newer.
AusBoss nodes 2.15.0 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.
The models:
minimax_h3_ref2va_pruned_int8_convrot.safetensors →
models/diffusion_models· 20.97 GB (the Ref2VA model, not FL2VA)qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors →
models/text_encoders· 15.69 GB (every H3 workflow uses it, so you may have it already)minimax_h3_video_vae_fp16.safetensors →
models/vae· 5.21 GB
Card older than the RTX 50 series? Get qwen3vl_32b_minimax_h3_int8_convrot.safetensors (27.14 GB) and pick it in Text encoder.
Only for Upscale 2x, SeedVR2:
seedvr2_3b_int8_convrot.safetensors →
models/diffusion_models· 3.46 GBseedvr2_ema_vae_fp16.safetensors →
models/vae· 501 MB
Optional, an uncensored text encoder (a community-made "Heretic" file; the regular one stays picked, see the second note in the workflow): qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors →
models/text_encoders· 15.68 GB
Tips
A view came out wrong? Run again. The seed changes after every run.
It is set up for photos. The prompt asks for a live-action photograph, and I only tested photos.
A phone or a camera in front of the face is fine. In my tests H3 left it out and drew the face.
Strong coloured light, like orange or red neon, becomes plain studio light, and in one test the skin came out lighter than in the photo. A photo in plain light keeps the person closest.
Each view's own picture is an exact quarter of the sheet, so a wide sleeve at the edge of a view can get cut off there. The sheet keeps it whole.
Going from half-body to full body can lose small things, like a shirt under a jacket. Name those in What the photo doesn't show too.
To change the views or the wording, open the box Write the sheet prompt. The prompt is plain text.
Times on an RTX 5090: about 45 seconds a sheet once the models are loaded, about 28 with Fast on.
LoRAs made for H3's Ref2VA model go in LoRA Loader.
Credits
MiniMax H3 is by MiniMax, under the MiniMax H3 Community License, and the files are Comfy-Org's repackage. The license only covers its "Applicable Territory", which leaves out the European Union, the United Kingdom, South Korea and the United States, and the same limit applies to what you make with it. Commercial products above US$20M a year in revenue need MiniMax's written OK and must show "MiniMax H3". SeedVR2 is by ByteDance Seed. The Heretic text encoder is Momoking's NVFP4 version of ethanfel's Heretic file, under Apache 2.0.
v1.0 (2026-10-26): first release.
Questions or bugs: GitHub issues or the comments here. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.