MiniMax H3 Image — Multi-Reference Edit Workflow & Use Cases (Patreon)
Downloads
Content
Hey everyone.
If you caught the public video ( https://www.patreon.com/aifuturetech/posts/minimax-h3-image-175033165 ), you saw what MiniMax H3 Image did on screen: text-to-image stress tests, the text-rendering poster, the mirror physics check, and the edits where the room stayed frozen and a person walked in.
This post hands you the workflow I use for the part that needs more setup: multi-reference editing, two character sheets plus a scene, one pass.
The two free templates from the video (t2i and single-reference edit) are also linked here so everything runs from one post. https://www.patreon.com/aifuturetech/posts/minimax-h3-image-175033165
WHAT TO DOWNLOAD (MODEL FILES)
Base model (pick ONE, from the Comfy-Org repo linked below):
- minimax_h3_image_pruned_int8_convrot.safetensors -> ComfyUI/models/diffusion_models/minimax-h3/
(INT8, my daily driver. The repo also ships BF16 and W4A8 if your VRAM wants otherwise.)
Text encoder:
- qwen3vl_32b_minimax_h3_int8_convrot.safetensors -> ComfyUI/models/text_encoders/minimax-h3/
(If you already ran the MiniMax H3 video model, skip this. The whole H3 series shares the same text encoder. The multi-ref workflow uses a heretic-tuned variant, qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors, same folder, if you want the exact graph from the video.)
VAE (image-specific, NOT the video one):
- minimax_h3_image_vae_fp16.safetensors -> ComfyUI/models/vae/
Optional prompt enhancer (my workflow keeps it switched off):
- qwen3.5_9b_int8_convrot.safetensors -> ComfyUI/models/text_encoders/
Model links, all of it in one place:
- Official model card: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
- Official PE repo + prompt formats: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image-PE/tree/main/prompt
- ComfyUI docs and templates: https://docs.comfy.org/tutorials/image/minimax/minimax-h3-image
- Quantized downloads (BF16 / INT8 / W4A8): https://huggingface.co/Comfy-Org/MiniMax-H3-Image/
QUICK START
1. Load h3_image_edit-multiReferences.json (attached). Update ComfyUI first, everything except the image comparer and the path loader is native nodes.
2. Check the three loaders point at the files above. No custom node hunt needed.
3. Wire your references into the conditioning in order. Picture 1 is the scene you are editing, later pictures are identity or style sources. Up to 16 images stack into the same conditioning.
4. Paste your prompt. Keep the prompt enhancer toggle OFF if your caption is already in the full <Picture>/<Subject> format.
5. Set the resolution in the ResolutionSelector. Leave the sampler alone. Queue it.
6. That's the whole loop.
2 USE CASES IN THIS WORKFLOW
1. Multi-reference scene build, the video's hero case.
Picture 1 is your environment, picture 2 is a character sheet, both stack into one conditioning and the graph puts that character into that room in one output. You can keep stacking toward the 16-reference ceiling, but 2 to 3 references per generation is the sane working zone.
2. Character sheet from one reference.
Got a single image of a character and want the turnaround? That same sheet is what use case 1 eats as picture 2, so it's the perfect prep step. Load the character as your only reference, leave the enhancer off, and paste this as the editing prompt:
"Convert the character from Picture 1 in the image to a Character Sheet showing a face close-up, front full body, side full body and back full body views."
One pass gives you the face close-up plus front, side and back full-body views, all locked to the same identity, white-background sheet style. From there you feed the sheet into use case 1 and the character walks into any scene you own.
WHAT TO INPUT (PROMPT FORMAT)
The edit prompt is a JSON. Four fields:
- user_editing_instruction: one short paragraph of intent. What changes, what must not.
- user_aspect_ratio_selection: "adaptive" or an explicit W:H like 16:9.
- expanded_caption: the long form. This is what actually drives the model.
- target_aspect_ratio: the resolved numeric ratio.
In the expanded caption, label every input once, then describe the result:
- "The living room in <Picture 1> is <Subject 1>, and the young woman shown in the character sheet of <Picture 2> is <Subject 2>."
- "The target image shows <Subject 1> with <Subject 2> sitting at the wooden desk, studying."
- Then a preservation clause per subject: "<Subject 1> is partially preserved" (list what stays exactly), "<Subject 2>'s face, hair bun, cream shirt and wristwatch are preserved from <Picture 2>."
- New props get no Subject number, just place them in the scene.
Short form is fine for simple swaps, one editing instruction with the rest blank, enhancer ON. Long form wins the moment identity matters, and you turn the enhancer OFF so the model reads YOUR caption, not a rewrite. Example pattern with two refs: scene as Picture 1, character sheet as Picture 2, one sentence naming each anchor you need to survive the edit.
KEY SETTINGS (AND WHY)
Steps: 8 on the first sampler. The model is distilled. Twenty steps overcooks the image; six to eight is the working window and I landed on 8 for the multi-ref graph.
CFG: 1.0. Distilled model, guidance is baked in. Nothing to tune.
Sampler: euler with the beta scheduler, denoise 1.0. Ships correct.
Resolution: start at 4MP (1:1 is 4096x4096 base at native). The conditioning node in the attached graph runs at 3104x3104 for a square edit. H3 Image gets mushy below that, which is why the graph upscales any low-res input before conditioning. Don't feed it 1024 and blame the model.
Negative conditioning: there is no negative prompt on this model. KSampler just leaves that socket unused. If the dead wire bugs you, the graph also shows the custom-sampler route: Basic Guider with one conditioning connection and the scheduler set separately. Performance difference is negligible, it's cleanliness, not quality.
Upscale pass (second sampler): multiply by 1.5, 3 steps, denoise 0.5, seed fixed at 42. That's the whole 4K tail. Any more steps and it starts over-sharpening and cooking noise. This is the section I'd call the find of the testing: three steps is enough.
Prompt enhancer toggle: OFF for finished captions, ON for plain English. One switch, no rewiring.
Output: 16-bit PNG, sRGB, through the save node into H3-Image/img.
----------------------------------------
TIPS
----------------------------------------
1. Reuse the H3 video text encoder if you own it. Same file, same folder, saves the download wait.
2. Write your long captions with an agent using the prompt folder in the official PE repo. It is just reference material for exactly this format, no model loading needed. That is how every caption in my video was made.
3. When editing a photo, name the small things you want kept, cables, pillow positions, window reflections. The model holds them when they're stated, and it's the difference between an edit and a regeneration.
4. Reflections change with lighting even when you didn't ask them to. That's the world-model side showing, but check the frame for it.
5. For multi-character, keep two reference sheets before you attempt more. The sixteen-reference ceiling is real but one generation per stacked scene is the sane pace.
6. Second-pass seeds: randomize the first sampler, fix the upscale pass. Consistent finishing, varied compositions.
----------------------------------------
COMMON ISSUES
----------------------------------------
Image looks cropped or wrong -> You're zoomed in on a 4096 canvas. Zoom out before judging the output.
Twenty steps, muddy overcooked look -> Back to 6-8 steps. This model is distilled, extra steps hurt.
Upscale pass is crispy with hot pixels -> Drop the second sampler to 3 steps at denoise 0.5. That's what the graph ships.
H3 nodes missing after loading -> ComfyUI is too old. Update, the model support is native now, check the update indicator mentions MiniMax sampling and VAE.
Edit changes the character's face -> Your caption probably has no preservation clause naming the anchors. One sentence, face plus hair plus outfit, per reference.
Out of VRAM -> Use the INT8 or the W4A8 build from the Comfy-Org repo, and keep the PE model unloaded if the toggle is off.
All JSON prompt examples from the video, every test, are in the attachment pile with this post. If something in the workflow behaves differently on your machine, drop a comment with your GPU, VRAM, and the exact error and I'll look at it.
Thanks for supporting the channel. You guys make these deep dives possible.
— Benji