MOSS-Image Character Sheet Designer for ComfyUI
Create character reference sheets with MOSS-Image — using text prompts or image-to-text analysis.
This project is based on my previous [H3 Character Sheet Designer](https://github.com/ukr8b3g-cmyk/H3-Character-Sheet-Designer) and [Qwen-Image 2.1 Character Sheet Designer](https://github.com/ukr8b3g-cmyk/Qwen-Image-2.1-Character-Sheet-Designer).
The interface and most features are similar, including illustrated view selection, character turnaround sheets, clothing details, part-specific prompts, styles, and automatic/manual layouts with up to 13 views.
## Important: This Is a Text-to-Image Workflow!
Unlike the H3 and Qwen versions, this MOSS-Image version does NOT directly use an input image as a visual reference for generation.
There are two input methods:
1. Text Prompt (Recommended)
Enter a description of your character and outfit directly. This is the primary workflow and generally gives you more control over the result.
2. Image Analysis
Load an image and enable enable_prompt_enhancer.
The workflow uses the MOSS text encoder's vision-language capabilities to analyze the image and convert its appearance into a text description.
Input Image → Image Analysis → Text Description → New T2I Generation
The image is not directly passed to MOSS as a reference image.
Even when starting from an image, the final character sheet is generated through text prompting — essentially T2I with an image-description step.
This means the generated character may look different from the original. Facial details, clothing, colors, accessories, and unseen angles may change.
Please do not expect exact character or image preservation. This is not image editing or reference-image conditioning.
## Prompting Tips
For the best starting point, describe one character and their appearance only.
Focus on the character's face, hairstyle, body proportions, clothing, colors, and accessories.
Avoid including background scenery, environments, unrelated objects, or multiple characters in your Character Profile prompt.
For example:
A female fantasy adventurer with short silver hair, green eyes, a dark leather jacket, brown trousers, and black boots.
Background descriptions or multiple people may confuse the character-sheet composition.
For complex layouts, the node can also reuse the connected MOSS text encoder to organize appearance details for individual views.
Results can improve with clear descriptions, but correct viewpoints, character consistency, and garment separation are not guaranteed.
## Required Models & Downloads
The supplied workflow uses the following files:
MOSS-Image SFT — Diffusion Model
[moss_image_sft_int8_convrot.safetensors]
or
MOSS_LUC [ moss-image-checkpoint ]
https://civarchive.com/models/3027480/mossluc-moss-image-checkpoint
Qwen3.5 9B — Text Encoder
[qwen3.5_9b_int8_convrot.safetensors]
VAE — FLUX.2 Small Decoder
[full_encoder_small_decoder.safetensors]
8-Step DMD LoRA — by lodestones
[lora_latest.safetensors]
(https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd/resolve/main/lora_latest.safetensors)
The example workflow starts with Euler / Simple, 8 steps, CFG 2 - 3, and LoRA strength 0.7- 0.8. These are starting settings, not guaranteed optimal values.
Set the CLIP loader type to moss_image.
## Installation
Open a terminal in your ComfyUI custom_nodes directory and run:
git clone https://github.com/ukr8b3g-cmyk/MOSS-Image-Character-Sheet-Designer.gitRestart ComfyUI and reload your browser after installation or updates.
A recent ComfyUI version with native MOSS-Image and subgraph support is required.
## GitHub & Workflow
Project Repository:
https://github.com/ukr8b3g-cmyk/MOSS-Image-Character-Sheet-Designer
Previous Character Sheet Designers:
H3: [https://github.com/ukr8b3g-cmyk/H3-Character-Sheet-Designer]
Qwen-Image 2.1: [https://github.com/ukr8b3g-cmyk/Qwen-Image-2.1-Character-Sheet-Designer]
For detailed instructions, model installation paths, known limitations, and updates, please see the GitHub README.
Remember: The MOSS version is primarily a prompt-driven character sheet generator. Image input is converted into text before T2I generation, not used for direct image-to-image reproduction.