H3 Image Character Sheet: one reference in, full character sheet out

Upload a picture of your character, pick a resolution, and press Run. The workflow renders 8 separate images with MiniMax H3 Image and stitches them into one character sheet.

What you get

• Top row (full body): front, left profile, right profile, back

• Bottom row (face close-ups): front, left profile, right profile, smiling

Each view is its own render, with its own prompt written for that angle. It works from your reference picture, so a photo stays a photo and an illustration stays an illustration. Markings, fur, hair, clothing and body proportions carry across all 8 views.

How to use it

1. Upload your character picture(s) in the Reference Images box. Extra pictures should be more angles of the same character.

2. (Optional) Describe the character in Character Notes. This helps with details the picture doesn’t show, like the back of a costume.

3. Set the Render Resolution. Every one of the 8 images is rendered at this size. The full-body views use the aspect ratio you choose (2:3 portrait by default), and the face close-ups are square at the same megapixels. You can use Aspect Ratio + Megapixels, or a Custom Width x Height.

4. Choose the Final Sheet Size from the presets:

   • 1920×1280

   • 2496×1664 (default)

   • 3072×2048

   • 3744×2496

   • 5024×3360

   • 3840×2160 (16:9)

5. Press Run.

The renders and the final sheet are set separately. You can render each image at high resolution for detail and still get a sheet of a sensible size. The renders are shrunk to fit and are never cropped or stretched.

Why it’s easy to use

• There are no Pose Studio or image-comparison parts, just upload, resolution and Run.

• Almost everything is tucked inside one Character Sheet Engine node. Double-click it to edit the 8 prompts, the sampler settings, or the model file names.

• The layout is organized into three panels: Upload, Resolution, Character Sheet.

Requirements

• ComfyUI v0.45.0 or newer. The MiniMax H3 Image Conditioning node isn’t in older versions.

• Custom nodes (all installable through ComfyUI-Manager):

  • ComfyUI-Pixaroma

  • rgthree-comfy

  • ComfyUI-Fantastic-MiniMaxH3-PromptBuilder, which provides the reference picture loader

• Models:

  • diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors (~21 GB), from Comfy-Org/MiniMax-H3-Image

  • text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors (~27 GB), from Comfy-Org/MiniMax-H3

  • vae/minimax_h3_image_vae_fp16.safetensors (~1.2 GB), from Comfy-Org/MiniMax-H3-Image

Tips

• Drafts: render at 4 MP per image. A full run at 4 MP per image took about 6 minutes on my setup.

• Final art: raise the Render Resolution for more detail. This is slower, so make a draft first.

• Rendering style: the face close-up prompts mention a “100mm portrait lens”, which suits photographic characters. For flat 2D art you may want to soften that wording in the Engine.

• Image content: the prompts assume one standing character with a face in the reference. Group shots, quadrupeds, or faceless designs may need the prompts adjusted.

• Left and right: “left” and “right” mean the direction the character faces in the image.

• Reroll: use the Seed box.

Credits

• Built on the Comfy-Org MiniMax H3 Image workflow, using the MiniMax H3 Image models.

• Inspired by foxydits H3 Image workflow https://civarchive.com/models/3045966/h3-image-w-pose-studio-other-advanced-features-foxfuressence

What’s new in v2.0

• Character Type dropdown: a native Custom Combo in its own pane under the uploader, with Character Notes beside it. It switches all 8 prompts to one of four sets. Each set is an editable column inside Engine › Shot Prompts.

• Four-legged and vehicle subjects: the body views render landscape automatically. Vehicles get front ¾, rear ¾, top-down and a detail close-up in the bottom row instead of faces. Landscape panels leave white bands above and below on a 3:2 sheet; that’s expected.

• Design first when nothing is uploaded: the workflow detects an empty upload and renders one design image, which then becomes Picture 1 for all 8 shots. With an upload, that extra render is skipped.

• Native nodes only: the new logic adds no custom node packs. It uses CreateList/GetItemFromList, StringFormat + StringCompare, a lazy If/Else Switch and RegexReplace.