H3 Character Sheet Generator
I started this since I was struggling to find good accurate high quality images I can use as reference for H3, so I realised this model has excellent character consistency and the ability to take up to 9 reference images.
Throw in some rough reference images, get back a character sheet you can reuse forever.
Like a lot of you I've been messing about with H3 and got a bit obsessed with the multi-image reference side of it. You can feed it up to 9 images, which means you can build a character out of bits — a face from here, armour from there, a hat from somewhere else — and it'll actually hold them together.
So this workflow does that, spins the character 360 (or 180 if using 4 panel version), and spits out a reference sheet you can use to keep them consistent in everything you make afterwards.
Why bother with a video model for this
If you generate six images of the same character separately, they'll disagree with each other. Jaw shifts, jacket changes, colours drift. You know the drill.
Six frames from one video generation can't do that — they come out of the same pass. That's the whole trick. The camera does a slow orbit with no hard cuts, the character stands still like a statue, and then the workflow grabs six frames and stitches them together.
How it works
- You drop in your images and describe them in the Input Text (A Prompt) box
- That added with a B Prompt which handles the spin, pose, lighting etc.
- It generates a slow 360 with no hard cuts, so the character stays consistent
- 6/4 frames get grabbed and stitched into the sheet
You also have the option for a 360 spin video and every individual frame as optional outputs, if you want to pick your own angles or use single frames as references later.
Two versions
| File | What |
|---|---|
H3_CharSheetMaker_6_Panel.json |
Front, both sides, back, plus two face shots |
H3_CharSheetMaker_4_Panel__Faster_.json |
Four views, ~40% fewer frames, noticeably quicker |
Same knight, 4-panel:
What you need
Models — every loader node has its download link baked in, so ComfyUI should just offer to grab whatever's missing when you open the workflow.
| Slot | File | Goes in |
|---|---|---|
| Diffusion | minimax_h3_ref2va_pruned_int8_convrot.safetensors |
models/diffusion_models/ |
| Text encoder | qwen3vl_32b_minimax_h3_int8_convrot.safetensors |
models/text_encoders/ |
| Video VAE | minimax_h3_video_vae_int8_convrot.safetensors |
models/vae/ |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors |
models/vae/ |
| Turbo LoRA (optional) | minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors |
models/loras/ |
Those are the INT8 low-VRAM builds. If you've got the headroom, higher precision versions give better detail — the text encoder especially, that's where prompt adherence lives.
Custom nodes — I tried to keep this close to vanilla. The only third-party stuff is in the optional Speed Ups group:
Delete that group if you want to.
Writing the A Prompt
One line per image. Say what to take, and — this bit matters more than people expect — say what to ignore:
<Picture 1> - keep the photo style. Use the bald head, facial hair, physique.
<Picture 2> - keep only the black outfit. Remove the hair, remove accessories.
<Picture 3> - use the shield. It is attached to the man's back.
If you don't rule things out by name, backgrounds and the wrong person's hair sneak through.
Also worth describing clothing in words, not just showing it. Faces carry across on their own but wardrobe drifts. "Black high-collared coat with silver buckles" sticks. "The coat from Picture 2" wanders off.
Anime to Real
Someone asked if it could do anime to realistic, so there's a modified B prompt for that too. It's surprisingly good at it.
Haruhi:
Sanji:
And here's the actual video that came from, so you can see how the thing works — slow spin, no cuts, then a couple of face shots at the end:
The six panels are just frames pulled out of that.
Objects
The model also does a excellent job of doing objects, however it is advised to show more angles as reference images.
Caveats (being honest here)
It's slooooow. You're generating 124 frames to use 6. That's the fundamental silliness of the approach. The 4-panel version helps by only generating 73 frames and taking 4 (40% speed up).
Speed-ups cost you something. Turbo LoRAs and caching do make it faster, but prompt adherence and quality take a slight hit. Worth it most times except final pass.
Quality is limited. It's a video model, it's better at video than stills. Crank the resolution and steps if you care, and pay for it in generation time.
The sheet alone might not be enough for close-ups. Resolution per panel is what it is. For close work I'd use the sheet plus some detail shots — face, clothing texture, whatever matters. And honestly, for a one-off video you might be better off skipping the sheet and just using your originals.
Tips
- More steps = slightly better quality, if you're willing to wait
- The B prompt says "neutral A pose" — delete that bit if you want a specific pose
- Use several different shots of the same character. More angles in, better 360 out
- Works on objects and props too. You'll want to tweak the B prompt, mainly swapping the face close-ups for detail shots
- Frame timings shift between runs. If a panel looks off, turn on the "save all frames" output, find a better frame number and plug it into the subgraph. Re-stitching is cached, so it's instant
Licence stuff
The workflow files are mine, do what you like with them.
The model isn't. MiniMax H3 has a community licence that excludes the EU, UK, South Korea and the USA, and it covers outputs as well as the weights. If you're planning anything commercial, go read the actual licence rather than taking my word for it. I'm not a lawyer.
Built on Comfy-Org/MiniMax-H3. Shout if you break it or improve it.
By C_Nugget
- Downloads last month
- -