H3 Character Sheet Creator
Turn 1–9 reference images into a character sheet in ComfyUI. Choose from eight editable layouts for face angles, expressions, full-body references, clothing details, or a custom portrait.
Download H3 Character Sheet Creator v1 · Download the original 10-example pack · GitHub and custom nodes
This repository contains the workflow, example images, prompts, and reproduction settings. Run generation in your own ComfyUI installation. Model weights are linked in the guide below.
Quick start
- Download the workflow JSON above and install the five custom-node packs listed in the guide.
- Select the H3 models, upload your reference image, and describe the character in BIND 1.
- Assign each reference a role in BIND 2.
- Preset 7 and 3:4 portrait are selected for the new layout. Choose another index 0–7 and set its canvas shape if desired, then run.
Recommended canvas ratios
| Index | Output layout | Recommended aspect ratio setting |
|---|---|---|
| 0 | Four faces + full body | 21:9 (Ultrawide) |
| 1 | Six faces + full body | 21:9 (Ultrawide) |
| 2 | Six headshots / expressions | 3:2 (Photo) |
| 3 | Four faces + neck-to-feet crop | 21:9 (Ultrawide) |
| 4 | Front + profile headshots | 3:2 (Photo) |
| 5 | Three columns: front + profile + body crop | 21:9 (Ultrawide) |
| 6 | Free text / single portrait | 3:4 (Portrait Standard); for the default portrait |
| 7 | Four heads above front crop + top-down + rear body | 3:4 (Portrait Standard) |
Set Canvas size → aspect_ratio to the value above after choosing a preset. Changing CHOOSE SHEET does not change the canvas automatically. Keep megapixels at 2.5 and multiple at 32 as a starting point. Preset 6 is editable free text: use 3:4 for its default single portrait, or choose a ratio that fits your custom composition. The saved workflow selects preset 7 and 3:4 portrait.
Preset 7 top row, left to right: front → three-quarter facing left → profile facing right → three-quarter facing right. Directions refer to the image edges. Below: front body cropped at the collar, overhead body, complete rear body. Use 3:4 (Portrait Standard).
New: preset 7 — four heads above front, top and rear body views
Lin — orbital pilot · 1408 × 1856 · sheet index 7 · seed 42
The top row reads left to right: front, three-quarter facing left, profile facing right, three-quarter facing right. Below are a front body crop from the neck to the feet, a top-down body view, and a complete rear view including the head. The canvas is 3:4 portrait.
Full-resolution sheet · Input portrait · Prompt
This unretouched sample uses the revised concise prompt at seed 42. Head angles can vary with the reference and seed.
This unretouched sample uses the public stock minimax_h3_ref2va_int8_convrot.safetensors overlay with the FL2VA base, 2.5 MP · 25 steps · ER-SDE · simple · CFG 1 · DMAD 0.8 · sigma shift 16/2. It reuses the original Lin portrait. To reproduce it, load that portrait into Picture 1, select face + hair, copy the character description from the linked prompt into BIND 1, and use preset 7 with the seed above. Per-sample model settings are recorded in the metadata.
Original ten examples
Five women and five men, with different ages, faces, hairstyles, builds, and wardrobes. These are original fictional characters. Each source portrait was generated with H3, then supplied to the character-sheet flow as a face + hair reference. The sheets are unretouched outputs.
The examples use 2.5 MP · 25 steps · ER-SDE · simple · CFG 1 · DMAD 0.8 · sigma shift 16/2. The existing batch used the locally prepared minimax_h3_ref2va_int8_convrot.safetensors.
Open a sheet for its full-resolution PNG. Each entry also includes its input portrait, prompt, and fixed seed. See all sample settings or the example index. The front body crops in layouts 3, 5 and 7 are intentional.
Five panels — four faces and a full-body view
Maya — travel photographer · 2464 × 1056 · seed 27101
Reference portrait · Full-resolution sheet · Prompt
Julian — city architect · 2464 × 1056 · seed 27106
Reference portrait · Full-resolution sheet · Prompt
Seven panels — angles, expressions, and full body
Lin — orbital pilot · 2464 × 1056 · seed 27102
Reference portrait · Full-resolution sheet · Prompt
Arjun — desert ranger · 2464 × 1056 · seed 27107
Reference portrait · Full-resolution sheet · Prompt
Six headshots — angles and expressions
Leila — archaeologist · 1984 × 1312 · seed 27103
Reference portrait · Full-resolution sheet · Prompt
Kenji — retired professor · 1984 × 1312 · seed 27108
Reference portrait · Full-resolution sheet · Prompt
Five panels — faces and clothing detail
Aoife — Victorian botanist · 2464 × 1056 · seed 27104
Reference portrait · Full-resolution sheet · Prompt
Two portraits — front and profile
Rowan — fantasy scholar · 1984 × 1312 · seed 27109
Reference portrait · Full-resolution sheet · Prompt
Three columns — front, profile, and body crop
Ines — ceramic artist · 2464 × 1056 · seed 27105
Reference portrait · Full-resolution sheet · Prompt
Mateo — starship mechanic · 2464 × 1056 · seed 27110
Reference portrait · Full-resolution sheet · Prompt
Installation and complete workflow guide
Install
Clone this repository into ComfyUI/custom_nodes/ComfyUI-H3-Character-Sheets. It registers only H3Subjects, H3RefSlots and H3ImagePromptCompose; it has no extra Python dependencies and does not download or load a VLM. Install the other four packs linked below, restart ComfyUI, then open the JSON in example_workflows.
cd ComfyUI/custom_nodes
git clone https://github.com/Jalen-Brunson/ComfyUI-H3-Character-Sheets.git
git clone https://github.com/ckinpdx/ComfyUI-MMH3Tools.git
git clone https://github.com/scottmudge/ComfyUI_MinimaxH3HybridLoader.git
git clone https://github.com/shootthesound/ComfyUI-Fizgig-H3-Still.git
git clone https://github.com/yolain/ComfyUI-Easy-Use.git
START HERE - H3 Character Sheets
One character, one generated sheet
- Install the packs listed in Dependencies and select your model files.
- Upload Picture 1. Describe the character in BIND 1.
- In BIND 2, match each active picture to its role.
- Set CHOOSE SHEET to 0-7. 7: four heads above front / top / rear body is selected by default.
- Set Canvas size → aspect_ratio from the preset table, then run. Inspect identity, profile, framing and clothing before keeping a sheet.
Read left to right. The eight editable sheet descriptions are below the main flow. “Single image” means one output image; you can supply 1-9 references.
Binding model - pictures are not subjects or output panels
Three different numbers
Subject 1 = the ONE character being depicted, even across many panels.
Picture 1-9 = the active uploaded reference images in list order.
Sheet index 0-7 = the desired output layout. It does not choose a reference.
H3Subjects → H3RefSlots → H3ImagePromptCompose
Identity + picture roles + ONE raw sheet description become ONE structured H3 prompt. The final prompt connects directly to conditioning and its preview.
Dependencies - all custom nodes
Install, then restart ComfyUI
- H3 Character Sheets: Subjects, Ref Slots, Image Prompt Compose.
- MMH3Tools: image list, reference conditioning, condition selector.
- H3 Hybrid Loader.
- Fizgig H3 Still: still latent + decode.
- Easy Use: Any Index Switch.
Use current ComfyUI with native H3, Resolution Selector and Text (Multiline). No private helper node, VLM model or separate still VAE is required.
Models - public downloads
Select the files installed on your machine
Comfy-Org / MiniMax-H3: FL2VA + Ref2VA diffusion models, H3 Qwen text encoder, video VAE and audio VAE.
Kijai / MiniMax-H3-experimental: DMAD 4-step full LoRA (rank 39 bf16).
The README has exact filenames and folders. No weights or personal reference images are included. Upload placeholders must be replaced before running.
Recipe
2.5 MP / 3:4 portrait (preset 7) / multiple 32; ER-SDE, simple, 5-25 steps, denoise 1; CFG 1; DMAD 0.8; sigma shift 16 / 2; hybrid AdaLN blocks 25-49; seed 42.
Select the downloadable stock Ref2VA overlay.
01 - choose and load reference images
Reference input checklist
Use clear, sharp pictures of the same character. A neutral front view is a useful primary face source; add a true profile if needed. Include an outfit/body reference when those details matter.
Picture 1 is enabled. Pictures 2-9 start bypassed. Upload the next picture, then right-click its node → Mode → Always (or Ctrl+B). Set its matching pic_N role in BIND 2.
Enable consecutive slots 1 through N. Bypassing or disconnecting a middle slot compresses the image list and renumbers later pictures. Move images and update their roles together.
Do not feed the same picture into two slots by accident. Each uploaded image is one picture even if it contains multiple panels.
When an input is already a character sheet
Use face sheet for a multi-view face-only reference; use character sheet for a sheet containing face, hair, outfit and body information. Use face (primary) for a single face photo where you want to lock that face as the primary.
02 - prompt binding instructions
BIND 1: who is being drawn?
Edit subject1_description: age range, face, hair, build and clothing. Keep Subjects 2 and 3 empty for this single-character workflow. All panels show Subject 1.
subject1_position is a textual alias, not panel placement. subject1_lora is only a trigger token: it does NOT load a LoRA. Leave it empty unless you add a character LoRA. The DMAD LoRA is a speed adapter, not an identity token.
Set setting to the desired backdrop. Leave loop_tail = latent_anchor and setting_as_subject = false. Extra rules are optional global constraints.
BIND 2: what should each picture contribute?
pic_N selects the role; pic_N_for = subject 1 selects the character. Match N to the active image-list order. Several pictures can belong to Subject 1.
Example: Picture 1 = face (primary); Picture 2 = hair; Picture 3 = outfit. Enable exactly those three loaders. Set every other pic_N to unused; leave all video/audio roles unused.
Roles, unused inputs, and diagnosis
face (primary): main face identity. face: supporting face angle. face + hair: both from one photo. hair: hairstyle. body: build/proportions only. outfit: clothes. appearance: overall look except face. face sheet / character sheet: a multi-view reference. setting: backdrop.
pic_N_note adds a qualifier such as “side profile”. Keep it short and factual.
unused removes a text binding; it does not disable the uploaded image. Also bypass the corresponding Load Image node, keeping active slots consecutive.
If the wrong face/outfit appears, compare the active image order with pic_N roles and inspect the final prompt. If panels are wrong, edit the selected raw template below. Never paste an already composed H3 prompt into a template.
03 - choose the output character sheet
Sheet presets and recommended canvas ratios
| Index | Output layout | Recommended aspect ratio setting |
|---|---|---|
| 0 | Four faces + full body | 21:9 (Ultrawide) |
| 1 | Six faces + full body | 21:9 (Ultrawide) |
| 2 | Six headshots / expressions | 3:2 (Photo) |
| 3 | Four faces + neck-to-feet crop | 21:9 (Ultrawide) |
| 4 | Front + profile headshots | 3:2 (Photo) |
| 5 | Three columns: front + profile + body crop | 21:9 (Ultrawide) |
| 6 | Free text / single portrait | 3:4 (Portrait Standard); for the default portrait |
| 7 | Four heads above front crop + top-down + rear body | 3:4 (Portrait Standard) |
Set Canvas size → aspect_ratio to the value above after choosing a preset. Changing CHOOSE SHEET does not change the canvas automatically. Keep megapixels at 2.5 and multiple at 32 as a starting point. Preset 6 is editable free text: use 3:4 for its default single portrait, or choose a ratio that fits your custom composition. The saved workflow selects preset 7 and 3:4 portrait.
Only the selected description is used. Edit its matching numbered text node below. Values 8-9 have no template; use 0-7.
04 - model setup
Put the two diffusion checkpoints in models/diffusion_models, the H3 encoder in models/text_encoders, and both VAEs in models/vae. Put DMAD in models/loras/minimax to match the saved selection, or select its installed path.
Hybrid: FL2VA base, Ref2VA overlay, block_range_adaln 25-49, final AdaLN from base. CLIP type must be minimax.
The video VAE serves both reference encoding and Fizgig decode. The audio VAE is connected because the H3 conditioning node requires it; no audio is saved.
Start with the saved sampler/LoRA settings. The “4step” filename does not require changing this saved still recipe to four steps.
05 - conditioning only
The selected prompt and ordered images enter H3 reference conditioning.
Keep total_frames = 5 and condition index 0. This node's latent output is intentionally unused: Fizgig Still Latent supplies the sampler's true one-frame latent.
Width and height come from the SAME Resolution Selector. Keep the multiple at 32. Never wire the conditioning node's clip latent into the sampler for this still workflow.
06 - sampling controls
The seed is fixed at 42 so comparisons are repeatable. Change it or choose randomize for variations.
25 steps / ER-SDE / simple / denoise 1 are inherited from the source. BasicGuider gives CFG 1. Fizgig batch_size = 1 produces one sheet.
More output pixels give each panel more room; more panels divide that same canvas into smaller views. Reduce reference count or resolution if memory is tight.
07 - choose a good generated sheet
Before reusing a result, check EVERY panel:
- Same face, hairline, skin tone, clothing and jewelry.
- Requested panel count and order; no duplicate figures.
- Profile is a true 90-degree view with the far eye hidden.
- Full hairstyle fits in headshots; hands and feet fit in body views.
- Similar head size and lighting across the headshots.
- The neck-to-feet crop is intentional only for options 3, 5 and 7.
Try another seed if a panel drifts. A simpler 2- or 3-view sheet can be more useful than a crowded sheet with inconsistent identity.
Output and prompt troubleshooting
PNG files are written beneath ComfyUI/output/h3_character_sheets/ and shown in Save Image. The saved PNG includes workflow metadata.
Inspect Final H3 prompt after a run. Expect one subject_definitions section and one detailed_description section, with Picture tags matching your active inputs.
Missing nodes: install all five packs and restart ComfyUI. Missing model: select an installed filename. Missing image: upload Picture 1 and every enabled optional image. Wrong layout: check sheet index 0-7 and the corresponding template.
Use plain text in descriptions. Avoid literal pipe characters: MMH3Tools treats a pipe as a separator between multiple prompts.
Exact model files
Download the following from Comfy-Org/MiniMax-H3, preserving the indicated ComfyUI model folders:
Folder under ComfyUI/models |
File |
|---|---|
diffusion_models |
minimax_h3_fl2va_int8_convrot.safetensors |
diffusion_models |
minimax_h3_ref2va_int8_convrot.safetensors |
text_encoders |
qwen3vl_32b_minimax_h3_int8_convrot.safetensors |
vae |
minimax_h3_video_vae_int8_convrot.safetensors |
vae |
minimax_h3_audio_vae_fp32.safetensors |
Download DMAD rank 39 bf16 into ComfyUI/models/loras/minimax/.
Source and credits
The installation and usage guide above follows the author’s edited GitHub README. The custom nodes and original workflow are maintained there. This page adds the ten-example gallery and downloadable sample bundle.
H3 models: Comfy-Org / MiniMax-H3. DMAD LoRA: Kijai / MiniMax-H3-experimental. Still-image latent and decoding: Fizgig H3 Still. The linked models and dependencies retain their respective terms. No model weights are redistributed here.
- Downloads last month
- -