MOSS-Image-Character-Sheet-Designer Introduction
MOSS-Image-Character-Sheet-Designer is a visual character-reference-sheet designer for ComfyUI. It helps you turn a character description into an organized sheet showing different views of the character, their clothing, and selected details.
Instead of writing a long layout prompt by hand, you can select illustrated view cards, choose a preset, and adjust the sheet through visual controls. The Designer supports up to 13 views:
- Six head views.
- Three full-body views.
- Front and back garment views.
- Hands.
- Footwear.
You can customize appearance details with Part prompts, add a global rendering style, and choose automatic or manual sheet dimensions. Version 1.0.1 also lets you edit the base instruction for each view.
This can help you explore a character concept, compare outfit details, or prepare a visual starting point for further drawing and design work. However, generated sheets still need review: the requested angles, panel count, anatomy, and clothing details are not guaranteed.
The author based this extension on H3 Character Sheet Designer and Qwen-Image 2.1 Character Sheet Designer, adapting the interface and prompt construction for MOSS-Image.
Important: this is a text-to-image workflow, even when you start from an image.
An uploaded image is first described in text. MOSS then generates a new sheet from that description. The image itself is not passed directly to the image generator as a visual reference. This is not direct image editing and does not guarantee preservation of the original character.
Download the workflow · Releases
How MOSS-Image-Character-Sheet-Designer Works
Think of the Designer as an art brief organizer. You describe one character, select the drawings you want on the sheet, and the Designer prepares instructions describing both the character and the arrangement.
1. Start with text or an image
Both input methods are available in the same workflow. You do not need separate templates.
Text input
Enter a description in Character Profile and set enable_prompt_enhancer to false. The workflow bypasses image captioning and uses your written description.
Describe one character and their appearance only, including:
- Face and hair.
- Body proportions.
- Clothing and footwear.
- Accessories.
- Important colors, materials, and distinctive details.
For example:
One adult character with a round face, short silver hair, and a compact build. They wear a navy cropped jacket with brass buttons, cream trousers, brown ankle boots, and a small red scarf.
Keep backgrounds, scenery, unrelated objects, and additional people out of Character Profile.
Image input
Choose your image in Load Image and set enable_prompt_enhancer to true. The workflow resizes the image to a maximum dimension of 1024 pixels before the existing MOSS text encoder describes it.
The process is:
Input image
↓
MOSS encoder analyzes the image
↓
A text description is produced
↓
The Designer prepares the sheet instructions
↓
MOSS generates a new image
This is similar to asking someone to describe a picture, then asking another artist to draw from that description without seeing the original.
Information can be lost along the way. Facial features, colors, garments, accessories, and lettering may change. Anything not visible in the image—such as the back of an outfit—must be inferred.
2. Choose the views and arrangement
Select illustrated cards or begin with a preset. The Designer uses your selection to prepare the panel arrangement and output dimensions.
- Auto scales the grouped layout within a selected megapixel budget.
- Manual retains the width and height you request.
The Layout area shows guide illustrations and panel positions. These are planning aids, not previews of the finished character.
3. Combine the instructions
The Designer combines the connected character description with relevant settings, including:
- The selected views.
- Each view’s base instruction.
- Applicable Part prompts.
- Global rendering style.
- Clothing shape and layout settings.
For certain selections, such as clothing views or three or more head views, the Designer can reuse the downstream MOSS encoder to extract part-specific features before preparing the final sheet prompt. Single-view generation and the working three-, five-, and seven-view presets bypass this internal analysis.
This feature extraction does not require a separate language model or external service.
4. Generate the sheet
For a single view, the Designer outputs a caption. For a multi-view sheet, it outputs MOSS Layout JSON—structured text describing the sheet’s content and placement.
You do not need to write this JSON yourself. The supplied workflow handles the connections to the image-generation nodes.
Layout instructions guide generation, but they do not act like rigid panel masks or a guaranteed collage of independently drawn images.
MOSS-Image-Character-Sheet-Designer Features
Illustrated view selection and presets
The view cards let you choose what appears on the sheet without memorizing special prompt syntax.
| View group | Available panels | Useful for |
|---|---|---|
| Head | Six head views | Exploring the face, hairstyle, and head shape from different directions |
| Full body | Three views | Reviewing the character’s overall silhouette and outfit |
| Garments | Front and back | Studying clothing separately from the wearer |
| Hands | Hands panel | Showing hand-related appearance details |
| Footwear | Footwear panel | Showing shoes or other footwear details |
Presets provide common selections that you can customize. The supplied workflow starts with five views.
More views are not always better. If a dense sheet loses detail, reduce the number of selected views or try a larger Auto budget.
Auto and Manual layouts
| Mode | Behavior | When to use it |
|---|---|---|
| Auto | Scales the grouped layout within 1, 1.5, 2, or 4 MP | When you want the Designer to manage the sheet size around your selected views |
| Manual | Keeps your requested width and height | When you need a specific output size or aspect ratio |
MP means megapixels: approximately one million pixels. A larger pixel budget gives a crowded sheet more room, but output size also affects resource use and rendering time.
For example, a small selection may be easier to read at a lower budget than an all-13-view sheet. If faces or garment details become too small, try 4 MP or split your exploration into smaller selections.
Clothing shape
Clothing shape offers three choices:
- Auto
- Pants
- Skirt or robe
Use these choices to guide the overall garment structure. For example, Pants requests a divided-leg silhouette, while Skirt or robe requests a continuous lower garment shape.
Clothing panels always request garments alone, laid flat. There is no Worn mode.
These are generation instructions, not automatic garment extraction or background removal. Body or neck fragments, changed clothing, and shape errors can still appear.
Part prompts
Part prompts add appearance details for selected parts without replacing the view’s basic instruction.
For example, you could add a garment detail such as:
Brass buttons, dark piping around the collar, and a small embroidered star on the left sleeve.
Use this area for focused appearance information rather than repeating the entire Character Profile. Keep the details compatible with your clothing-shape and other settings.
Global style
The Style control adds a rendering style across the sheet. The supplied workflow starts with no added global style.
A global style helps you request a shared visual treatment across the selected panels. It does not guarantee identical details between views.
Editable view prompts
In version 1.0.1, the Prompts tab appears after Settings. It lets you replace the built-in base instruction for any of the 13 views.
To edit a view:
- Open Prompts.
- Select the view you want to change.
- Edit its instruction.
- Select Apply.
- Save the workflow to retain the change.
Use Reset this prompt to restore only the selected view’s built-in instruction.
The distinction between the two prompt controls is important:
| Control | What it changes |
|---|---|
| Part prompts | Adds appearance details |
| Prompts | Replaces a selected view’s base instruction |
An explicitly empty view instruction is allowed. It clears only that base instruction; the connected character description, applicable Part prompts, and settings are still added.
Each view instruction accepts up to 2,000 UTF-16 code units. Most common characters count as one unit; an emoji typically counts as two.
Custom instructions are saved per node and restored when you reload the workflow. They also remain present after encoder feature analysis in both Caption and Layout JSON output.
The settings preview is not the final generation prompt. It does not include the connected external description or the final encoder-processed runtime text.
Saved settings and language
The interface language follows your ComfyUI language setting.
View choices, layout settings, styles, and prompt edits are saved with each Designer node. The Settings area exposes the saved JSON for recovery. Invalid saved JSON is reported rather than silently replaced.
Before opening a workflow with custom view prompts in an older extension version, keep a backup. Older versions may not understand the optional view_prompts field.
Using MOSS-Image-Character-Sheet-Designer
Open the supplied Text or Image workflow. Its model loading, image description, and sampling are grouped inside a subgraph, a collapsed group of connected nodes. The Designer remains outside for easier access.
A simple text-based session
- Set
enable_prompt_enhancerto false. - Enter one character and their outfit in Character Profile.
- Choose a preset or select individual view cards.
- Adjust Clothing shape if needed.
- Add optional Part prompts or a global style.
- Choose Auto or Manual sizing.
- Set the seed and run the workflow.
- Review the resulting viewpoints, clothing, anatomy, and panel count.
A simple image-based session
- Select your own file in Load Image. The default filename is only a placeholder.
- Set
enable_prompt_enhancerto true. - Keep the image-description instructions focused on the subject and outfit.
- Choose views and optional design settings.
- Set the seed and run.
The Generate Text prompt inside the subgraph can include a requested change. That change is handled through text description and fresh generation, not by editing the original pixels.
Starting generation settings
The supplied workflow uses:
| Setting | Template default |
|---|---|
| Views | Five |
| Global style | No added style |
| Sampler | Euler |
| Scheduler | Simple |
| Steps | 8 |
| CFG | 2 |
| LoRA strength | 0.8 |
A seed controls the starting randomness. Keep it fixed when comparing a setting change, or change it to explore another result.
CFG controls how strongly generation follows its guidance. The author suggests optionally comparing CFG 3, which improved some tested clothing details but did not guarantee correct viewpoints.
More steps add rendering time. These settings are starting points, not universal quality recommendations.
MOSS-Image-Character-Sheet-Designer Models
The supplied workflow uses one MOSS-Image-SFT configuration made from four components. These components work together; they are not four alternative image-generation modes.
| Component | File | Role |
|---|---|---|
| Diffusion model | moss_image_sft_int8_convrot.safetensors | Generates the image from the prepared instructions |
| Text encoder / CLIP | qwen3.5_9b_int8_convrot.safetensors | Processes text and supplies image understanding for Generate Text |
| VAE | full_encoder_small_decoder.safetensors | Converts the generated internal representation into a visible image |
| 8-step LoRA | lora_latest.safetensors | Adapts MOSS-Image-SFT for the community eight-step generation approach |
The CLIP loader type must be moss_image. The tested encoder already provides the vision capability needed by Generate Text; no separate CLIP Vision loader is needed.
The 8-step LoRA’s model card recommends strength 0.7, while the supplied Designer template uses 0.8. These are distinct starting settings. Consult the LoRA model card when comparing settings.
Model weights are not included or downloaded automatically.
MOSS-Image-Base and MOSS-Image-SFT
The upstream MOSS-Image project provides:
- MOSS-Image-Base: the base text-to-image model.
- MOSS-Image-SFT: a version fine-tuned from Base on curated data.
The upstream project demonstrates more organized layout and text grouping with SFT in one comparison. This is not a character-sheet quality guarantee.
The Designer template specifically uses SFT with the eight-step LoRA. The supplied documentation does not establish an alternative Base configuration or a controlled Base-versus-SFT character-sheet comparison.
What’s New with MOSS-Image-Character-Sheet-Designer
Version 1.0.1 — Editable view prompts
This update gives you more direct control over how each panel is requested.
- New Prompts tab: edit the base instruction for any of the 13 views.
- Per-view Apply and Reset: customize one instruction without resetting the others.
- Empty overrides supported: clear a base instruction while retaining the other applicable prompt content.
- Persistent edits: custom instructions are saved per node and restored on reload.
- Analysis preservation: edits remain present after feature analysis in Caption and Layout JSON output.
- Per-instruction limit: each instruction accepts up to 2,000 UTF-16 code units.
Existing workflow files, node identifiers, sockets, model selections, and sampler settings remain unchanged. Older workflows retain their original defaults until you edit them.
The package version is 1.0.1, but the node display name still includes v1.0. The internal prompt compiler remains moss-sheet-1.4.1; this is a separate identifier, not a conflicting package version.
Validation and what it means
For version 1.0.1, the author reports:
- 24 Python tests passed.
- 9 JavaScript/DOM tests passed.
- Chrome save/reload restoration and the live preview API were checked locally.
- Prompt preservation after feature analysis was tested with controlled feature JSON, including empty outfit facts.
The local checks used ComfyUI 0.44.0, frontend 1.57.6, and Python 3.12.12. The original workflow was also tested with ComfyUI 0.43.0 and frontend 1.57.4.
No new encoder/GPU generation or visual-quality benchmark was run for 1.0.1. The update validates prompt-editing behavior, not improved image quality or guaranteed compliance with custom instructions. Physical Windows IME and Nodes 2.0 acceptance are not claimed.
See the CHANGELOG and validation documentation.
Troubleshooting MOSS-Image-Character-Sheet-Designer
The generated character does not match my input image
This is a limitation of the image-to-text-to-image process.
- Confirm that
enable_prompt_enhanceris true. - Check that Load Image contains the intended file.
- Make essential appearance details explicit in the description instructions.
- Try text input when you need precise control over which details are described.
- Compare another seed.
Exact identity, lettering, garment construction, and unseen surfaces are not guaranteed. If direct visual-reference conditioning is essential, review the separate H3 and Qwen designers linked below; they use a different approach.
Views are missing, repeated, or facing the wrong direction
- Reduce the selected view count.
- Try 4 MP if a dense layout lacks detail.
- Change the seed.
- If you edited a view’s base instruction, use Reset this prompt and compare again.
- Keep additional instructions focused and compatible with the requested view.
Exact profiles, three-quarter views, head counts, and all-13 layouts can still vary.
Clothing panels include a neck, body fragments, or changed garments
- Check that Clothing shape matches the intended outfit.
- Describe the garment clearly in the character description and applicable Part prompts.
- Reset custom garment-view instructions if they conflict with garment-only output.
- Try another seed or fewer selected panels.
- Optionally compare CFG 2 and CFG 3.
Clothing panels request garments alone, laid flat, but the workflow does not perform guaranteed segmentation or extraction.
Background content appears unexpectedly
- Remove scenery, locations, unrelated objects, and extra people from Character Profile.
- Keep the description focused on the subject and outfit.
- Review any background-related settings separately.
A background choice is a generation instruction, not a segmentation mask. The Designer does not remove backgrounds or guarantee that conflicting background instructions are ignored.
Error: E_PROMPT_EMPTY
- For text input, set
enable_prompt_enhancerto false. - Enter a nonempty Character Profile and ensure it reaches the Designer’s
promptinput. - For image input, choose a valid image and enable the image branch.
- Run again.
Error: E_ANALYSIS_ENCODER
- Check that the Designer’s
promptoutput connects directly to one positive CLIPTextEncode. - Confirm that this encoder uses the MOSS CLIP.
- Set the CLIP loader type to
moss_image. - Keep negative instructions in a separate CLIPTextEncode.
The Designer reuses the downstream positive encoder for part analysis. It has no visible CLIP input and no negative output.
Feature analysis reports invalid JSON
- Read the reported error.
- Review the connected character description for irrelevant or conflicting content.
- Simplify the description if necessary.
- Retry the workflow.
Invalid feature JSON fails explicitly. There are no hidden corrective retries.
My prompt is rejected for length or Unicode
- Shorten the external character description to no more than 2,000 UTF-16 code units.
- Check any edited view instruction against its own 2,000-unit limit.
- If invalid Unicode is reported, retype or replace the affected text.
This is the extension’s input limit, not MOSS-Image’s token limit. Overlong input is rejected rather than silently cut off.
Models or nodes are missing
- Check that the workflow’s loaders select files available in your environment.
- Compare the filenames with the Models section.
- Confirm that ComfyUI supports MOSS-Image and native subgraphs.
- Update ComfyUI if core nodes such as Generate Text, Switch, or Seed are unavailable.
- Restart ComfyUI and reload the browser after updates.
Generation takes longer than expected
Image captioning, part analysis, model loading, and image rendering are separate stages.
- Use text mode to bypass image captioning when an image is unnecessary.
- Start with fewer views.
- Avoid increasing steps unless you are deliberately testing them.
- Compare seed changes without changing the description or layout when possible; valid cached analysis may be reused.
The author measured warmed image captioning alone at roughly 3–6 seconds locally, with about 6.3 seconds for the selected instruction. This excludes model loading, internal analysis, and diffusion rendering. It is not a total generation-time estimate, and performance varies by hardware.
Frequently asked questions
Do I need separate workflows for text and image input?
No. Both methods are included in the same workflow and selected with enable_prompt_enhancer.
Does an empty view prompt disable that panel?
No. It clears only the panel’s base instruction. Deselect the view card to remove the panel.
Does the settings preview show everything sent to MOSS?
No. It excludes the connected external prompt and encoder analysis.
Do I need another language model or an online service for part analysis?
No. The Designer reuses the MOSS encoder already connected downstream.
Can I use negative prompts?
Keep them in a separate CLIPTextEncode. The Designer does not provide a negative output.
Why does an older saved description seem to be ignored?
Old character/outfit fields are retained for compatibility, but the connected prompt input is authoritative.
What should I check when opening an older workflow?
Review the connected prompt and negative-prompt wiring. Legacy clothing_mode is removed during normalization, and old negative-output links must be moved to a separate encoder. Connected links are not silently deleted.
MOSS-Image-Character-Sheet-Designer Inputs and Outputs
You normally do not need to change the supplied connections. This table is useful when checking a modified workflow.
| Connection or saved field | Meaning |
|---|---|
prompt input | Required external text description of the subject and outfit |
state_json | Saved view, layout, style, and interface settings |
prompt output | A single-view caption or MOSS Layout JSON |
width and height outputs | Output dimensions for EmptyFlux2LatentImage |
Connect the prompt output directly to the positive CLIPTextEncode, and connect the dimension outputs to EmptyFlux2LatentImage.
There is no duplicate internal character-description box in the Designer: the connected external description supplies that content.
Learn More about MOSS-Image-Character-Sheet-Designer
Designer resources
- Text or Image workflow — the supplied starting workflow.
- Releases — downloadable versions and release information.
- CHANGELOG — recorded extension changes.
- Testing documentation — the scope of validation and its limitations.
MOSS-Image guidance
- MOSS-Image project — background, model information, and examples.
- English MOSS-Image user guide — prompting, layout formats, and inference recommendations. Its general settings are separate from this extension’s eight-step template defaults.
- Comfy-Org MOSS-Image model collection — ComfyUI-format model files.
- 8-step LoRA model card — adapter-specific recommendations and checkpoints.
- Browser-based MOSS layout editor — a separate community tool for exploring layout prompts visually. It makes no image-generation calls and is not required by this extension.
Related character-sheet designers
- H3 Character Sheet Designer — uses native image-reference conditioning rather than converting the input image into text.
- Qwen-Image 2.1 Character Sheet Designer — a separate reference-based designer with its own models and controls.
These projects share design ideas, but their settings, models, and release histories are not interchangeable with the MOSS extension.
General ComfyUI learning
- ComfyUI documentation — explanations of workflows, nodes, and creative tools.
- ComfyUI workflow library — examples for learning visual workflows.
Useful ComfyUI shortcuts include Ctrl+Enter to queue a generation, Ctrl+S to save, and Ctrl+O to open a workflow. On macOS, Cmd can replace Ctrl.
MOSS-Image-Character-Sheet-Designer License
The extension’s original code and documentation use the scoped MIT License.
Artwork and third-party materials have separate rights described in NOTICE.md. Model weights retain their respective licenses.
The repository’s license does not grant model rights or generated-image rights. Review the relevant model terms and the rights associated with any reference images you use.
