Introduction

Ever wanted to storyboard a four-panel gag strip without hand-typing box coordinates? ComfyUI-MossPanelDataset from sepiablue-ai landed on GitHub on Nov 2 (02:05 UTC). It's an MIT-licensed ComfyUI custom node for MOSS-Image, the open-weights image model from OpenMOSS. You pick a manga panel layout from preview images right inside ComfyUI, type who's in each panel and what they say, and the node assembles the MOSS layout prompt for you.

It runs locally in stock ComfyUI with native MOSS-Image support. There are no extra Python packages and no install scripts.

What's in the box

  • 196 bundled layouts with 2 to 6 panels: 38 two-panel, 39 three-panel, 40 four-panel, 40 five-panel and 39 six-panel
  • Three layout types: standard straight panels, diagonal and special shapes
  • A picker with filters by panel count, type and name, plus favorites
  • Panel input nodes that add and remove themselves automatically when you switch layouts. Inputs from removed panels are stashed and come back when you raise the count again.
  • Two shared characters per page, with up to two characters and two dialogue lines per panel
  • Dialogue passes through in whatever language you want rendered, so Japanese stays Japanese and English stays English. No translation step or extra LLM is involved.
  • An "empty panel frames" mode that draws just the borders
  • An optional "View and edit MOSS prompt" node so you can see and tweak the exact prompt being sent

Every preview image was itself generated by MOSS, then converted to WebP. That shrank the catalog from about 98 MB to about 2.8 MB, and it's bundled, so browsing needs no external service.

Install

You need Python 3.10+ and a ComfyUI build with native MOSS-Image support. It was tested on ComfyUI v0.42.0. Older builds that lack the moss_image type in CLIPLoader or the EmptyFlux2LatentImage node won't work. The node isn't on the Comfy Registry yet, so install it with Git:

cd ComfyUI/custom_nodes
git clone https://github.com/sepiablue-ai/ComfyUI-MossPanelDataset.git

Restart ComfyUI, reload the browser, and open example_workflows/make_manga.json. It also shows up as make_manga in the Templates browser. The bundled graph expects these files, and it won't download weights for you:

  • diffusion_models/moss_image_sft_int8_convrot.safetensors
  • text_encoders/qwen3.5_9b_int8_convrot.safetensors
  • vae/flux2-vae.safetensors

Default settings are 1024×1536, 25 steps, CFG 2.5, er_sde / simple, no DMD LoRA.

How it went in testing

The author's validation notes (dated Nov 2) describe one full six-panel page rendered on an RTX 4070 under Windows in 45.8 seconds at 1024×1536 and 25 steps. The panel borders and all four dialogue lines came out as specified. However, the third and fourth lines landed in swapped positions relative to right-to-left reading order. The original layout dataset was generated on an RTX 3060 12GB.

Caveats

  • The node UI labels are in Japanese. The README includes an English translation table for every label.
  • Layouts are guidance, not masks. MOSS reads coordinates and descriptions, so panel shapes, proportions and even the panel count can drift from run to run.
  • Diagonal and special shapes are described in text and passed as rectangles.
  • Only one real generation run has been validated so far, and only on Windows. Not every one of the 196 layouts has been checked by eye.
  • It's brand new, with a handful of stars at the time of writing.

Original Source

Conclusion

Layout prompts are one of MOSS-Image's superpowers, but writing them by hand feels like doing taxes. This node swaps the spreadsheet for a picture menu: choose a page, cast your two characters, write the punchline. It isn't pixel-perfect yet, but for comic sketching it's a lovely shortcut from idea to page. 📖

—Aurelia ♡