CivArchive
    H3 Character Sheet Designer for ComfyUI -Image Generation - v1.0
    NSFW
    Preview 144650199
    Preview 144650003
    Preview 144650004

    H3 Character Sheet Designer for ComfyUI

    v1.5.0

    Both MiniMax H3 Image and P-H3 now display v1.5.0. The Head · 6 views preset uses the same three-column, two-row layout and screen-facing directions with Layout Image on or off. In No clothing · natural body surface mode, six pale areas mark the positions in the guide, while the prompts specify each view’s direction. Generated angles can vary with the model, reference image, and seed.

    v1.5.0 improves left/right guidance for Head 6, especially with Layout Image enabled. A generated view may still face the opposite way.

    Download v1.5.0 — node package and both workflows · Workflow JSONs only · GitHub

    Stop ComfyUI and run the command from your ComfyUI folder.

    New installation

    git clone https://github.com/ukr8b3g-cmyk/H3-Character-Sheet-Designer.git custom_nodes/H3-Character-Sheet-Designer

    Update an existing Git installation

    git -C custom_nodes/H3-Character-Sheet-Designer pull --ff-only

    Restart ComfyUI, refresh your browser, and open either v1.5.0 JSON from the Designer’s workflows folder. MiniMax H3 Image requires ComfyUI 0.45.0 or later.

    v1.4.0

    Dedicated MiniMax H3 Image support with output up to 16 MP.

    The new workflow uses MiniMax H3 Image Conditioning and opens with Standard 5 views and High detail. Auto also offers 8 MP and 16 MP.

    The ZIP includes the new MiniMax H3 Image v1.4.0 workflow and the previous P-H3 v1.3.0 workflow. Each uses its own conditioning node and model set.

    The new workflow requires H3 Character Sheet Designer v1.4.0 or later and ComfyUI 0.45.0 or later. GitHub Release

    MiniMax H3 Image Model Downloads

    Diffusion model
    Save in ComfyUI/models/diffusion_models/minimax/
    minimax_h3_image_pruned_int8_convrot.safetensors

    Text encoder
    Save in ComfyUI/models/text_encoders/
    qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

    VAE
    Save in ComfyUI/models/vae/
    minimax_h3_image_vae_fp16.safetensors

    What’s New — v1.3

    • The new v1.3 workflow uses Pika H3 Flex Ref2VA LoRA.

    • MOSS-style presets, including 7 views with hands and feet on the right.

    • One shared layout guide for ON/OFF, with smaller margins and larger full-body guides.

    • Cleaner UI and minor clothing/13-view prompt refinements. Major quality improvements are not confirmed; human remnants and missing views can still occur.

    • Includes v1.3 and the previous v1.2.2 workflows. Auto sizing may increase rendering time and VRAM usage.

    • Models and downloads

      ComponentDownload / filenameFolder under models/Diffusion / templateDaSiWa Hybrid V4 INT8 — DasiwaMinimaxH3_dasiwaHybridV4_int8.safetensorsdiffusion_models/minimax/Native alternativeH3 Ref2VA INT8 — minimax_h3_ref2va_pruned_int8_convrot.safetensorsdiffusion_models/Text encoder / CLIPH3 Qwen3-VL INT8 — qwen3vl_32b_minimax_h3_int8_convrot.safetensorstext_encoders/Alternative encoderH3 NVFP4 AWQ — qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorstext_encoders/Video VAEH3 VAE INT8 — minimax_h3_video_vae_int8_convrot.safetensorsvae/Alternative VAEH3 VAE FP16 — minimax_h3_video_vae_fp16.safetensorsvae/Pika LoRA / enabledRef2VA — pika_h3_flex_ref2va_v1_bf16.safetensorsloras/minimax/Pika LoRA / disabledFL2VA — pika_h3_flex_fl2va_v1_bf16.safetensorsloras/minimax/

    v1.2

    13 Views, 4 Reference Images & External Seed Control

    • Up to 13 views, including head, full-body, clothing, hands and footwear.

    • One character image + three optional references of the same character.

    • Fixed / randomize Seed controls outside the subgraph.

    • Previous nodes and workflows remain available.

    Update the custom node before loading the new workflow. No character-sheet LoRA is required; the current template requires rgthree-comfy.


    V1.0

    Create character sheets from a reference image using MiniMax H3’s built-in reference conditioning and prompt understanding. Select the views you need, adjust the output size, and optionally add English prompts for clothing or other details.

    No character-sheet LoRA is required. This Designer is the only custom node in the included workflow; generation uses ComfyUI Core nodes. You’ll need a compatible ComfyUI build and the H3 diffusion model, text encoder, and video VAE. Models are not included.

    Installation

    Open a terminal in your ComfyUI/custom_nodes folder and run:

    git clone https://github.com/ukr8b3g-cmyk/H3-Character-Sheet-Designer.git

    Restart ComfyUI and refresh the browser page. Download the official workflow, load it into ComfyUI, and select your reference image and models.

    English Part prompts are recommended. Hand and footwear close-ups are optional; start without them unless needed. Results may require prompt adjustments. To generate a single view, select only that view and queue manually.

    A simple, intuitive character-sheet designer using the MiniMax H3 models available through ComfyUI’s official templates. FL2VA can work well for straightforward character sheets. If you want to change clothing, colors, or other details, consider trying the official Ref2VA model from Comfy-Org. A Hybrid model is also linked in the workflow notes as an optional alternative. Results vary by model, reference image, and prompt.

    For model links, examples, and full instructions, see the GitHub README.

    Description

    The included workflow is a minimal example to help you get started. It does not use Turbo LoRAs or other optional acceleration techniques. Feel free to customize it for your setup, including adding compatible LoRAs or other nodes inside the subgraph. It is intended as a starting point, rather than a fully optimized workflow.

    FAQ

    Comments (15)

    wyxzddsjj919Oct 5, 2026· 2 reactions
    CivitAI

    这个帅,能为QW21或者K2 EDIT制作1个版本吗 ?

    nekotxtOct 5, 2026
    CivitAI

    It works well, and I was surprised to be able to get a large output without oom. Thanks for this wf!

    NoDroneOct 5, 2026
    CivitAI

    This is cool thx, but would it be possible to use with a GGUF version of minimax H3?

    ukr8b3g201
    Author
    Oct 6, 2026· 1 reaction

    The usual GGUF custom node is ComfyUI-GGUF. For MiniMax H3, however, please use an H3-compatible version recommended by the model provider. For example, leejet’s H3 GGUF models specify leejet’s ComfyUI-GGUF fork.

    Open the H3 subgraph, replace “Load Diffusion Model” with “Unet Loader (GGUF),” select the corresponding H3 Ref2VA GGUF, and reconnect its MODEL output to the same nodes. The Designer itself shouldn’t need any changes, but I haven’t tested this GGUF setup myself.

    NoDroneOct 6, 2026· 3 reactions

    @ukr8b3g201 I will try that thank you.

    Update**: Just did what you said. Went into the subgraph and replaced the "Load Diffusion Model" with "Unet Loader (Dynamic Vram)" then re-connected the nodes .... works perfectly. This is great, thanks so much!

    ukr8b3g201
    Author
    Oct 6, 2026· 1 reaction

    @NoDrone Thanks for testing it and reporting back! Good to know it works with “Unet Loader (Dynamic Vram).” That’s helpful information for others wanting to use GGUF too.

    SERGEY73Oct 8, 2026· 2 reactions
    CivitAI

    How many frames should I set in the MiniMax H3 node?

    By default, the Reference to Video node shows a custom / NaN error, and the Image from Batch node shows the same error.

    What is the recommended number of frames for this workflow?

    ukr8b3g201
    Author
    Oct 15, 2026

    Sorry for the late reply! For this character-sheet workflow, set length to 5 frames in the MiniMax H3 Reference to Video node. In Image from Batch, set batch_index to 0 and length to 1 to extract the first frame.

    These are the settings used in the official template. If you still see custom / NaN after entering these values, please share a screenshot of both nodes and your ComfyUI and frontend versions so I can check further.

    Jojo0GojoOct 10, 2026· 2 reactions
    CivitAI

    Nice workflow and impressive node, thank you !

    Would be great if more reference images could used, separated as clothes, head, etc. and additional layouts for head.

    ukr8b3g201
    Author
    Oct 15, 2026· 1 reaction

    Thanks for the feedback, and sorry for the late reply! I’ve added support for multiple reference images. This update was made with requests like yours in mind, so you can use separate references for clothing, the head, and other details. Thanks for the suggestion!

    Jojo0GojoOct 15, 2026

    @ukr8b3g201 Sounds great, additional head angles, which provide a 360° view, with like 4 shots, would be amazing !

    mehjabin1000cmcolaOct 16, 2026

    一张图里放多个参考的话,脸部的像素降低,一致性会不会差呢

    ukr8b3g201
    Author
    Oct 16, 2026· 1 reaction

    @mehjabin1000cmcola Thanks for the question! In the H3 version, the lower layout preview isn’t passed directly to the model as a reference image. That part is guided by text, so consistent facial identity across views isn’t guaranteed. The results can also look a little blurry sometimes.

    The H3 and Qwen versions work differently internally. I deliberately chose this approach for H3 because I personally preferred the results to using the layout image directly. If facial consistency is your priority, you might prefer trying the Qwen version:
    https://civitai.red/models/2986253/qwen-image-21-character-sheet-designer-image-generation

    mehjabin1000cmcolaOct 16, 2026· 2 reactions

    @ukr8b3g201 Thanks♪(・ω・)ノ!!

    Workflows
    MiniMax H3

    Details

    Downloads
    810
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/4/2026
    Updated
    11/20/2026
    Deleted
    -

    Files

    H3_Character_Sheet_Designer_wf.json

    Mirrors

    h3CharacterSheetDesigner_v10.zip

    Mirrors