Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-MossPanelDataset

ComfyUI Extension: ComfyUI-MossPanelDataset

Repo Name

ComfyUI-MossPanelDataset

Author
sepiablue-ai (Account age: 467 days)
Nodes
View all nodes(6)
Latest Updated
2026-11-05
Github Stars
0.04K

How to Install ComfyUI-MossPanelDataset

Install this extension via the ComfyUI Manager by searching for ComfyUI-MossPanelDataset
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-MossPanelDataset in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-MossPanelDataset Description

ComfyUI-MossPanelDataset prepares manga layouts and structured prompts for MOSS-Image, combining panel selection, character descriptions, camera compositions, and dialogue in ComfyUI.

ComfyUI-MossPanelDataset Introduction

ComfyUI-MossPanelDataset is a custom-node extension that helps you create multi-panel manga pages with MOSS-Image inside ComfyUI. You choose a page layout from preview images, describe your characters and scenes, and enter dialogue in dedicated fields. The extension then prepares the structured prompt that MOSS needs to generate the page.

Think of it as a visual storyboard assistant: you decide what happens in each panel, while the extension handles much of the layout information and prompt assembly.

You do not need to write coordinates or JSON by hand for normal use, and you do not need a separate workflow for every page layout.

The panel-layout picker inside ComfyUI

What you can create

  • Manga pages containing 2–6 panels, selected from 196 bundled layouts.
  • Stories with up to three shared characters.
  • Panels containing up to three characters and three dialogue entries.
  • Silent panels, background-only panels, or pages containing only empty panel frames.
  • Pages guided by 120 camera-composition previews.
  • Characters described in your own words or through appearance text extracted from reference images. The extension is particularly useful when you want to try different page arrangements without rebuilding your workflow, keep character descriptions together, or plan a conversation panel by panel.

Important: This is a prompt-and-layout guidance tool, not a page compositor. MOSS generates the complete image. The extension does not paste panels together, overlay finished lettering, or repair borders and text after generation.

Watch the basic workflow

Basic manga-generation demo — MP4, 30 seconds

The recording shows choosing a four-panel layout, entering characters and dialogue, and generating a page. Waiting time is shortened, and Japanese operation captions appear at the top. The manga image itself is not edited.

Alternative hosted video link

How ComfyUI-MossPanelDataset Works

ComfyUI uses connected blocks called nodes. A workflow is the collection of nodes that describes how your image will be made. This extension adds nodes for page layouts, shared characters, individual panels, compositions, appearance descriptions, and optional prompt editing.

1. Choose the page structure

A layout controls the overall arrangement of panels on the page.

For example, you might choose a four-panel arrangement for a short conversation, then switch to a six-panel arrangement to add more story beats. The extension automatically adjusts the associated panel-content nodes and their connections.

The preview is a guide, not an image that is copied into your result.

2. Create a shared character sheet

You enter each character’s name and appearance once in Shared characters. Individual panels then refer to those characters.

For example:

  • Name: Mina
  • Appearance: A young adult woman with a short black bob, round glasses, and a loose yellow cardigan. The name is for your own reference and can be in any language. Appearance descriptions should be in English.

This shared setup avoids retyping the same appearance description for every panel, although exact visual consistency is still not guaranteed.

3. Describe each story beat

Each Panel content node describes one panel:

  • Where the scene takes place.

  • Which characters appear.

  • What happens.

  • Who speaks and what they say. For example:

  • Scene: Inside a quiet café. Mina looks surprised as her friend places a small gift on the table.

  • Characters: Character 1 and Character 2.

  • Speaker 1: Character 1. - Line 1: You remembered! Scenes should be written in English. Dialogue stays in the language you want MOSS to render.

4. Optionally guide the camera

A composition controls the view inside a panel: a close-up, full-body shot, high-angle view, over-the-shoulder view, or another arrangement.

A useful distinction is:

| Setting | What it controls | | --- | --- | | Layout | How panels are arranged across the page. | | Composition | How characters and the camera are arranged inside one panel. | | Scene | What is happening, including emotions, actions, props, and lighting. |

The extension adds the composition’s English description to the panel scene and scales its character-placement regions to fit the panel.

5. Build the MOSS prompt

MOSS-Image supports structured layout prompts that describe both content and positions. The extension combines your selected layout, character descriptions, scenes, dialogue, and optional compositions into this structured format.

You can think of the result as a storyboard brief with approximate placement boxes attached.

These boxes guide the model; they do not force it to obey an exact outline. This is why generated panel shapes, character positions, and lettering may differ from the previews.

What happens when you use character images?

Character-image support follows a separate preparation step:

Reference image → Qwen appearance description → shared character settings → MOSS generation

The reference image itself is not passed to the image-generation model. Qwen describes the visible character in English, and MOSS uses that description.

Normal text-based prompt assembly does not require translation or an additional language model. Optional image-to-appearance extraction reuses the Qwen model already used by MOSS.

ComfyUI-MossPanelDataset Features

Visual layout catalog

The bundled catalog contains 196 layouts, covering two through six panels.

| Panels per page | Available layouts | | --- | ---: | | 2 | 38 | | 3 | 39 | | 4 | 40 | | 5 | 40 | | 6 | 39 | | Total | 196 |

You can filter layouts by:

  • Panel count.
  • Type.
  • Name.
  • Favorites. The default view shows standard straight layouts. You can switch to diagonal or special layouts when you want a less conventional page structure.

For example, you can begin with a straight four-panel page to establish the story, then explore diagonal layouts for a more dynamic arrangement. Special shapes remain model guidance rather than guaranteed geometry.

The layout data and previews are bundled, so browsing the catalog does not require an external service or a runtime database connection.

Automatic panel-node management

When you choose a layout with a different panel count, the extension adds or removes the associated panel-content nodes and connections.

Inputs from removed panels are temporarily stored. If you return to a larger panel count, those inputs are restored. Character, dialogue, model, sampler, and seed settings are preserved.

This makes it easier to test page pacing. For example, you can compare a six-panel version with a four-panel version without immediately losing the text from the extra panels.

Automatic adjustment works with the extension’s layout node and MossMangaPanel. A panel-content node that is also connected to other nodes is not automatically deleted.

Save your workflow to preserve your work, including stored inputs from removed panels.

Shared characters and dialogue

You can define up to three shared characters and choose any supported combination of them for each panel.

To enable Character 3, fill in both:

  • Character 3 name.
  • Character 3 appearance. If both fields are empty, the workflow continues to use the two-character setup.

Each panel supports up to three speaker-and-line pairs. A speaker must also be included among that panel’s characters.

| Desired panel | How to configure it | | --- | --- | | One character speaking | Select that character, then choose them as the speaker and enter the line. | | A silent reaction | Select the characters, choose No dialogue, and leave the dialogue text empty. | | Background only | Select No characters and describe the environment in the scene field. | | Three-character conversation | Include all three characters and use up to three speaker-and-line pairs. |

Enter dialogue as plain text, without surrounding quotation brackets or JSON.

Speech-balloon regions use a simple fixed arrangement:

  • One line: upper center toward the right.
  • Two lines: upper right and upper left.
  • Three lines: upper right, upper center, and upper left. The extension does not analyze the scene to find unused space or determine balloon-tail directions.

Camera-composition catalog

The Choose a composition node provides 120 previews:

  • 40 for one character.
  • 40 for two characters.
  • 40 for three characters. You can search or filter by character count, category, name, English prompt, and favorites.

To use a composition:

  1. Open a composition-enabled sample such as composition_duo.json.
  2. Select a preview in the composition node.
  3. Connect its Composition for the panel output to the Composition input of the relevant Panel content node.
  4. Make sure the composition’s character count matches the panel’s selected characters.
  5. Describe the scene and dialogue, then generate. Selected characters are assigned to composition positions in numerical order. Changing who speaks first does not swap character positions.

Composition descriptions specify the camera, shot size, character placement, body and face direction, pose, and furniture used for staging. They deliberately leave expressions, emotions, conversational actions, props, and lighting to your scene description.

For example, you can keep an across-the-desk composition while changing the scene from a friendly meeting to a tense argument.

The same composition can guide tall or wide panels, but an exact match to its square preview is not guaranteed.

Character appearance from images

The Character appearance (image / text) node offers three methods:

| Method | What happens | When it is useful | | --- | --- | --- | | Text only | Passes your appearance description through unchanged. Image and Qwen processing are skipped, even if connected. | You already have a description or want direct control. | | Extract from image | Uses Qwen to turn visible appearance into English text. | You want to begin with an existing character reference. | | Extract from image + text | Appends your supplementary description to the extracted text. | You want to add details or requirements alongside the image description. |

Extraction describes features such as apparent age, facial features, build, hairstyle, clothing, and accessories. A separate short age-group question produces a prefix such as:

Apparent age group: young adult

Supplementary text is not translated, so write it in English for consistency with the appearance prompt.

Using your own reference

  1. Open composition_character_images.json.
  2. Select your images in the Load Image nodes. Reference images are not bundled.
  3. Choose the extraction method for each character.
  4. Run the workflow and inspect the appearance text shown in the node.
  5. If needed, press Copy result to the text field and switch to text only, then refine the description. When building your own workflow, connect the appearance node’s Appearance for shared characters output to the relevant character appearance input. Convert the appearance field to an input if necessary.

Each appearance node accepts one image. A reference containing a single person is recommended. For multi-person images, the extraction instruction focuses mainly on the large foreground person.

For analysis only, larger images are proportionally reduced to a maximum long side of 768 pixels. Your original image is not modified.

Reusing extracted descriptions

Once you are happy with the text, you can copy it into text-only mode, remove the upstream reference-image nodes, and save the workflow. This lets you reuse the character description elsewhere without relying on those image files.

Normal ComfyUI caching is used. Changing only the composition or seed does not rerun appearance extraction.

Extraction is descriptive, not an identity-locking system. It may misjudge age, infer unseen clothing, or include art-style details. Review the text before relying on it.

Character-image demo — MP4, 30 seconds Alternative hosted video link

The demo shows loading two references, choosing compositions and a page layout, and generating the page. Waiting time is sped up; the final manga image is not edited.

Empty panel-frame generation

Set Output on the layout node to Empty panel frames to omit character, scene, and dialogue content from the prompt.

This is useful when you want MOSS to generate a blank page structure rather than a finished story page. Because MOSS still draws the frames, their exact geometry and count are not guaranteed.

Optional prompt viewing and editing

Use make_manga_edit_prompt.json when you want to inspect or adjust the exact structured prompt.

The View and edit MOSS prompt node sits between the layout node’s MOSS prompt output and CLIPTextEncode.

  1. Set Prompt to use to As received and run the workflow.
  2. Inspect Received prompt. If the edit box is empty, it is populated with the same text.
  3. Edit the prompt.
  4. Switch to Edited text and run again to use your changes. After changing characters or dialogue upstream, switch back to As received, or use Copy received prompt to the edit box to refresh the editable version.

Your manual edits are never overwritten automatically.

The edited prompt must remain valid JSON. If a bracket or comma is broken, execution stops and reports the error’s line and column. You do not need this feature for ordinary manga creation.

Language support, favorites, and saved work

  • When ComfyUI’s display language is Japanese, the extension’s interface appears in Japanese.
  • For other ComfyUI display languages, the interface appears in English.
  • Workflows saved in either interface language can be reopened without conversion.
  • Input-validation messages explain the affected field and correction in both Japanese and English. Your characters, dialogue, settings, and stored panel inputs belong in your saved ComfyUI workflow.

Favorites are different: they are stored in your browser’s local storage. They do not automatically transfer to another browser and are removed if you clear the site’s data. The extension does not write your changes into its bundled catalog JSON files.

ComfyUI-MossPanelDataset Models

The extension does not include its own image-generation model or download model weights automatically. Its bundled workflows use MOSS-Image and the following three components.

Models used by the bundled workflows

| Component | File | Role | | --- | --- | --- | | MOSS image-generation model | moss_image_sft_int8_convrot.safetensors | Generates the manga page from the assembled prompt. | | Qwen text encoder | qwen3.5_9b_int8_convrot.safetensors | Encodes prompts for MOSS and is reused for optional character-image appearance extraction. | | FLUX.2 VAE | flux2-vae.safetensors | Converts the generated internal image representation into the visible image. |

In the workflow, CLIPLoader must use type=moss_image.

See the Comfy-Org MOSS-Image model page and the official MOSS ComfyUI guide for model-specific details.

MOSS-Image Base and SFT

The wider MOSS-Image project provides two main model variants:

| Model | Description | Relevance here | | --- | --- | --- | | MOSS-Image-Base | The base text-to-image model. | Useful for learning about the underlying model family. The bundled extension workflow does not select it by default. | | MOSS-Image-SFT | Fine-tuned from Base on curated data. | The bundled workflows use an SFT-format model distributed for ComfyUI. |

The official MOSS comparison shows clearer text grouping and a more organized whiteboard layout with SFT in that example. This is not a guarantee of perfect manga lettering or panel arrangement.

The provided documentation does not establish extension-specific performance comparisons for alternative model variants or community adapters. Start with the bundled workflow configuration when learning the extension.

Tested environment and workflow settings

The author reports testing with ComfyUI v0.42.0 and Python 3.13. Python 3.10 or later and native MOSS-Image support are required.

The original layout dataset was generated on Windows with an RTX 3060 12GB. This is a reported generation environment, not a universal minimum-hardware specification. Actual generation on other GPU and operating-system combinations was not verified in the extension documentation.

The bundled workflow uses:

| Setting | Value | | --- | --- | | Image size | 1024 × 1536 | | Steps | 25 | | CFG | 2.5 | | Sampler | er_sde | | Scheduler | simple | | DMD LoRA | Not used |

These are the extension’s sample settings. They differ from the standalone MOSS project’s general defaults, so use the matching workflow as your starting point.

ComfyUI-MossPanelDataset Sample Workflows

For your first page, open make_manga.json. You can drag it into ComfyUI or open make_manga under ComfyUI-MossPanelDataset in the Templates browser.

Choose the sample that matches your goal

| Workflow | What it demonstrates | Best starting point for | | --- | --- | --- | | make_manga.json | A four-panel page with two shared characters, using text descriptions and no composition nodes. | Learning the basic layout, character, scene, and dialogue controls. | | make_manga_edit_prompt.json | Basic manga generation with a prompt-viewing and editing node. | Inspecting the generated prompt or making direct JSON edits. | | composition_duo.json | A four-panel conversation with two characters and preset compositions: face-to-face, side by side, across a desk, and a close-up argument. | Planning camera views and the relationship between two characters. | | composition_1to3.json | Three panels containing one, two, and three characters, with matching compositions and up to three dialogue entries. | Learning mixed cast sizes and third-character settings. | | composition_character_images.json | Three appearance-input methods combined with composition-guided generation. | Starting character descriptions from your own reference images. | | composition_duo.api.json | An API-format version of the two-character composition workflow. | Automated API use rather than editing nodes in the normal interface. |

The five non-API workflow files can be opened by dragging them into ComfyUI.

The character-image sample does not include reference images. Its text-only Character 3 demonstrates configuration and does not appear in the default panels.

A simple first-page process

  1. Open make_manga.json.
  2. Press Choose a layout from images and select a preview.
  3. Enter your shared character names and English appearance descriptions.
  4. Write each panel’s scene in English.
  5. Select the characters present in each panel.
  6. Choose speakers and enter dialogue in the language you want rendered.
  7. For silent panels, choose No dialogue and leave the line empty.
  8. Press Run in ComfyUI.
  9. Save the workflow to keep your story and settings. Useful ComfyUI shortcuts include Ctrl + Enter to queue generation and Ctrl + S to save a workflow. On macOS, ComfyUI also supports replacing Ctrl with Cmd.

What’s New with ComfyUI-MossPanelDataset

The supplied documentation identifies the following version-specific and additional changes. Changes without an explicit version number are kept separate rather than assigned to an invented release.

Version 0.1.1: bundled JSON layout data

All 196 layouts are read from data/layouts.json.

Why it matters: The distributed extension does not need a database connection at runtime. The experimental SQLite database and generation history are not included, while layout IDs, names, arrangements, compositions, and previews remain unchanged.

Expanded composition catalog

The composition catalog now contains 120 templates, with 40 each for one, two, and three characters. The original 20 two-character templates retain their IDs and names.

Why it matters: You have more camera and staging options without losing selections based on those original templates.

Composition-only prompt cleanup

The author revised 46 composition prompts that previously included expressions, emotions, conversational actions, props, or lighting. Their previews were regenerated with MOSS.

IDs, names, and character regions remain unchanged.

Why it matters: Composition templates leave more creative decisions to your scene text. Saved workflows still open, but the changed prompt wording can produce different images even with the same seed.

Support for three shared characters

You can use a third shared character and specify up to three characters and three dialogue entries per panel.

Why it matters: You can move between solo reactions, two-person conversations, and group scenes within the same page.

Appearance extraction from reference images

Character appearance can be supplied as text, extracted from an image, or extracted and supplemented with text.

Why it matters: You can begin with an existing visual reference without adding another model specifically for extraction.

Age-aware character proportions

Earlier page prompts applied natural adult proportions to everyone. The current prompt asks MOSS to follow the age and body proportions in each character’s appearance description.

Why it matters: Character descriptions can guide a wider range of apparent ages and builds. Results may differ from older workflows even with the same seed, particularly when the appearance text does not specify age.

Preview-size optimization

Layout and composition previews use quality-90 lossy WebP files:

  • Layout previews total approximately 2.8 MB.
  • The 120 composition previews were reduced from approximately 178 MB to 25 MB. The images were converted without added lettering, compositing, cropping, resizing, or border redrawing. Fine details may differ from the original PNGs because of compression.

Distribution status and updating

At the time described in the supplied README, Registry version 0.1.0 was flagged and 0.1.1 had not yet been published. Check the Registry public API for the current status.

After updating the extension, restart ComfyUI and reload the browser page. Keep your own workflows saved separately; your story content is not stored in the bundled catalog files.

Planned work is listed in ROADMAP.md.

Troubleshooting ComfyUI-MossPanelDataset

Symptoms: CLIPLoader has no moss_image type, or EmptyFlux2LatentImage is unavailable.

  1. Check whether your ComfyUI build includes native MOSS-Image support.
  2. Update ComfyUI if those components are missing.
  3. Restart ComfyUI and reload the browser page.
  4. Reopen the sample workflow and verify the loader selections. See the official MOSS ComfyUI guide for the supported workflow.

The workflow cannot find a model file

  1. Inspect the image-model, text-encoder, and VAE loader dropdowns.
  2. Compare their selections with the filenames in the Models section.
  3. If your files have different names or are in subfolders, select the actual files in the loaders.
  4. Confirm that CLIPLoader uses moss_image. The extension does not automatically download missing weights.

A speaker-validation message stops generation

  1. Open the panel named in the error.
  2. Check its selected characters.
  3. Include the intended speaker in that panel, or change the speaker to a character who is present.
  4. For Character 3, confirm that both the shared name and appearance fields are filled.
  5. For a silent panel, select No dialogue and clear the line text. Validation messages identify the affected field and explain the correction in English and Japanese.

Changes to dialogue are not appearing in the result

If you are using the prompt-editing workflow:

  1. Open View and edit MOSS prompt.
  2. Check whether Prompt to use is set to Edited text.
  3. Switch to As received to use the latest automatically assembled prompt.
  4. Alternatively, copy the received prompt into the edit box, then reapply your edits. The edit box intentionally keeps your manual changes instead of updating itself automatically.

The edited prompt reports a JSON error

  1. Note the line and column in the error message.
  2. Check that location for missing commas, quotation marks, or mismatched brackets.
  3. If you do not need manual editing, switch to As received.
  4. To start over, copy the received prompt into the edit box. This issue applies to direct prompt editing; ordinary scene and dialogue fields do not require JSON.

A composition gives unexpected character positions

  1. Make sure the composition is connected to the intended panel.
  2. Match the preview’s character count to the panel’s cast.
  3. Remember that characters are assigned in numerical order.
  4. Do not expect changing dialogue order to swap their positions.
  5. Check your scene description for instructions that conflict with the selected composition. The preview remains guidance rather than an exact pose or placement constraint.

Image extraction describes the wrong person, age, or outfit

  1. Use a reference containing one clearly visible character.
  2. Read the extracted text shown in the appearance node.
  3. Choose Copy result to the text field and switch to text only.
  4. Correct the age group, clothing, body proportions, or unwanted art-style wording.
  5. Generate again using the revised description. Extraction can infer details that are not visible. Manga faces can also appear younger than the intended age, so age classification may be inaccurate.

The character-image sample cannot load its references

  1. Open each Load Image node.
  2. Choose your own character image.
  3. Run the workflow again. The filenames in the sample refer to testing images; the images themselves are not included.

Removed panels do not disappear, or an old layout is unavailable

  • A panel node connected to other nodes is intentionally protected from automatic deletion.
  • Automatic adjustment is designed for the extension’s layout node and MossMangaPanel.
  • Older prototype-only or excluded layouts are not part of the distributed 196-layout catalog. Select a current layout for those workflows.
  • After an update, restart ComfyUI and reload the browser.

Favorites disappeared

Favorites are browser-local, not part of your workflow file.

  1. Check whether you changed browser or browser profile.
  2. Check whether site data was cleared.
  3. Recreate the favorites in the browser you now use. There is no documented automatic transfer between browsers.

Frequently asked questions

Why does my result have a different panel count or border shape? MOSS follows descriptions and rectangular regions rather than a rigid page mask. It can change the count, proportions, or shape of panels. Diagonal and special layouts are not enforced with polygon masks.

Can it guarantee correct dialogue lettering? No. Misspellings, distorted text, and incorrect balloon placement can occur. The extension does not repair lettering or overlay corrected text afterward.

Does a reference image preserve exact character identity? No. It is converted into a text description, not supplied directly as an image reference to the generation model.

Can I use more than three characters or three dialogue entries per panel? No. The documented controls support a maximum of three shared characters, three characters per panel, and three dialogue entries per panel.

Does it translate my dialogue? No. Dialogue is passed through in the language you enter. Character appearance and scene descriptions should be written in English.

Does browsing the catalog require internet access? No. Layout data and previews are bundled. Browsing them does not require an external service.

Why did the same seed produce a different result after an update? A seed is only one part of generation. Revised composition wording or age-and-proportion instructions can change the output even when the seed remains the same.

Are the previews guaranteed examples of accurate layouts? No. They are MOSS-generated examples. Panel counts and near-duplicate candidates were checked automatically; the author does not claim that every layout received a complete human review for distinctness or accuracy.

Learn More about ComfyUI-MossPanelDataset

Extension resources

MOSS-Image guidance

General ComfyUI learning

  • ComfyUI documentation — reference material for understanding workflows and nodes.
  • ComfyUI workflow library — examples for learning the wider ComfyUI environment. These are general resources, not guaranteed compatible replacements for this extension’s samples.

ComfyUI-MossPanelDataset License and Sources

The extension’s original code, layout data, and bundled preview images are distributed under the MIT License. See LICENSE for the full terms.

The previews were generated with MOSS-Image. The supplied documentation reports that the official MOSS-Image code and Comfy-Org model weights were published under Apache License 2.0, checked on November 2, 2026.

Model weights, ComfyUI, and third-party implementations are not included in this repository and are not covered by the extension’s MIT license. Check the relevant distribution’s license when using those components.

ComfyUI-MossPanelDataset Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.