Generating Consistent Multi-Angle Character Sheets in Google Colab with MiniMax-H3 and P-H3


Eight-angles full-body character sheet (3872 x 2176) generated with P-H3

The H3 Character Sheet Generator notebook can be used to generate character sheets with different compositions and layouts from one or more reference image(s). It supports both MiniMax-H3-Image and P-H3-Image models, allows you to mix and match up to 8 head angles, 8 full-body angles, and isolated detail panels, and includes built-in spatial layout guides, optional face restoration, and automatic panel cropping. Shown below are examples of 8 MP and 4 MP character sheets generated from a single image using the notebook.


Reference Image 1:

Character Sheet (Custom selection of 2 head and 8 body angles (4320 x 1856)):


Reference Image 2: 

Character Sheet (Custom selection of 4 head and body angles (2688 x 1536)):


Reference Image 3: 


Character Sheet (3 Head and Body angles (2048 x 2048)):



Reference Image 4: 
Character Sheet ( 6 Head and 3 Body angles (3168 x 1280)):



Reference Image 5:

Character Sheet ( 7 Views (2 Head + 3 Body + Hands & Feet) (3168 x 1280)):




What You Need Before Starting

  • Hardware: A Google Colab GPU runtime (tested on an L4 GPU with High-RAM enabled).
  • Reference Images: 1 to 3 clear images of your subject and their accesories. A well-lit front portrait or full-body image works best as your primary identity anchor.

Usage Guide

The workflow is organized into three sequential cells. Once your environment and reference images are loaded in the first two cells, you only need to re-run the third cell whenever you want to generate new variations.

1. Prepare Environment & Download Models

The first cell installs the required dependencies and downloads the required models. 

  • Model Suite Selection: Use model_download_selection to download MiniMax-H3-Image Only, P-H3-Image Only, or Both. I have not tested all the generation options provided by the notebook but so far P-H3-Image has generated better results than MiniMax-H3-Image.

  • Presets & Custom URLs: Both suites offer quantized and full-precision presets (INT8 ConvRot, W4A8, BF16, and GGUF). Pasting a direct HuggingFace or Civitai URL into any custom_*_url field automatically overrides the dropdown preset for that specific component.

  • Optional LoRAs: You can enable and download up to five custom LoRAs directly into your session.
  • The retry_face_model_download checkbox can be enabled and the cell run again to ensure the face restoration models are downloaded in the event that they failed to download the first time the cell was run. If one of the models fail to download, then the generation cell will output an error message during face restoration.

2. Upload & View Media Bank

Rather than re-uploading files every time you tweak a prompt or angle, the second cell stores your reference images in a persistent Media Bank inside /content/ComfyUI/input.

  1. Run the cell and select one or more images from your device (or click Cancel upload if your files are already in the bank).
  2. Review the numbered thumbnail grid (Image [1], Image [2], etc.) and note the index numbers you want to reference in the generator cell.

3. Configure & Generate Your Character Sheet

The third cell brings all creative and technical controls together into five clean form sections:

Section 1 — Media Bank Reference Selection: Enter your primary character's index number in ref_image_1_index. You can optionally attach up to two supplementary references (such as an alternate angle, outfit, footwear, or prop) by entering their index numbers in ref_image_2_index and ref_image_3_index, or leave them at 0 to skip.

Section 2 — Presets & View Checkboxes: Choose one of the built-in turnaround presets (such as 3 Head + 3 Body, 6 Head + 3 Body, or 8 Head + 8 Body Full 360°), or leave sheet_preset on Custom (Use Checkboxes Below) to toggle individual head angles, full-body angles, hands, feet, and isolated flat-outfit panels. Enable apply_face_restoration to automatically refine all detected faces. Enable use_face_setting_2 to ensure that all faces in the character sheet are restored; disable it if it gives weird results like adding faces on angles where the face is slightly visible. If apply_face_restoration is enabled, then both the sheet with bad faces and that with the restored faces will be displayed. But if only_show_face_restored_sheet is enabled, then only the sheet with restored faces will be displayed at the end of the generation. Face restoration is mainly for realistic characters and may or may not work for other styles.

Section 3 — Layout, Conditioning & Resolution: Switching conditioning_node between MiniMax H3 Image Conditioning and P-H3 Image Conditioning automatically loads the matching model files and sets the optimal step count, CFG, sampler, and scheduler. Choose between Stacked Rows, Columns (MOSS / H3 Designer), or Uniform Grid layouts, toggle the spatial layout guide (with or without visible panel grid lines), set your target resolution (2.5 MP, 4.0 MP, or 8.0 MP), and enable auto_crop_panels to export each view as its own individual PNG. The cropping only works properly when there's enough space between the images in the character sheet. Disabling the  spatial layout guide sometimes results in the character sheet not adhering to the settings in Section 2.

Section 4 — Style & Part Prompts: Keep art_style on Match Reference Style to preserve the exact medium and lighting of your input image, or select a dedicated production style. Fine-tune expressions, clothing overrides, and hand/foot details using the dedicated part prompt fields.

Section 5 — Seed & Sampler Overrides: Change seeds when the results are not satisfactory. Leave override_sampler_settings unchecked to use the automatic model defaults, or check it to manually adjust steps, CFG, sampler, scheduler, and sigma shift. If you downloaded a LoRA in cell 1, then ensure that the LoRA strength is not equal to zero.


Test Results & Performance

Generations on the L4 GPU usually take an average of 3 minutes 30 seconds, and this depends on the options enabled or selected in the generation cell. P-H3-Image has better prompt-adherence and gives better results than minimax-H3-Image overall. Some of the views sometimes fail to generate properly, but changing or randomizing the seed can improve resuts. 

Comments

Popular posts from this blog

How to use Wan2.2 Google Colab Notebooks for Image to Video Generation

Generating Long, High-Quality AI Videos with LongCat-Video in ComfyUI (Google Colab Setup)

I dumped Kling AI for these colab notebooks