MOSS Manga Colorist v1

A community-trained MOSS-Image Base LoRA for black-and-white manga colorization, with luminous lighting, warm skin tones, and cinematic anime-style rendering.

Give it a manga page as the reference image and use the prompt comic. The examples below use LoRA strength 1.0, 30 steps, CFG 3, and seed 7. All six LoRA examples are the first output generated for their page, with no rerolls or retouching.

Download the LoRA | Inference config | Training details | Example manifest

Colorized samples

Six first-pass LoRA outputs using the prompt comic. See the comparisons below for each original page and the base-model result.

Prompt
comic
Prompt
comic
Prompt
comic
Prompt
comic
Prompt
comic
Prompt
comic

Example gallery

Every comparison uses the same column order: Original | Base Model (No LoRA) | Trained LoRA (First Pass). Click a comparison to open the full-size image; the links beneath it provide the original and generated PNGs.

These are comparisons of two practical workflows. The base model uses native image-to-image sampling and a detailed colorization prompt written for each page. The LoRA uses its trained reference-conditioning workflow and the fixed prompt comic. The workflows also differ in precision, step count, and resolution, so these images do not isolate the effect of switching LoRA on or off.

Gintama - Clean Shonen Linework

Gintama: original, base-model colorization, and trained LoRA colorization

Original PNG | Base-model PNG | LoRA PNG | Settings

Solanin - Detailed Slice-of-Life Art

Solanin: original, base-model colorization, and trained LoRA colorization

Original PNG | Base-model PNG | LoRA PNG | Settings

JoJo's Bizarre Adventure - Bold Lines and Sculpted Figures

JoJo's Bizarre Adventure: original, base-model colorization, and trained LoRA colorization

Original PNG | Base-model PNG | LoRA PNG | Settings

A Silent Voice - Delicate Coming-of-Age Art

A Silent Voice: original, base-model colorization, and trained LoRA colorization

Original PNG | Base-model PNG | LoRA PNG | Settings

Maison Ikkoku - Classic Romantic Comedy

Maison Ikkoku: original, base-model colorization, and trained LoRA colorization

Original PNG | Base-model PNG | LoRA PNG | Settings

Hellsing - Heavy Blacks and Angular Shading

Hellsing: original, base-model colorization, and trained LoRA colorization

Original PNG | Base-model PNG | LoRA PNG | Settings

How to use

This adapter was validated with the MOSS image-to-image path in anima_lora, revision 83b8d6d582dd. Set up that repository using its installation guide, then download the MOSS Base model pack through its GUI or make download-model ARGS=moss-base.

From the installed anima_lora directory, download this release:

uv run --no-sync hf download moonrailgun/MOSS-MangaColorist-v1 --local-dir manga-colorist

Generate the first example using the included config:

uv run --no-sync python scripts/moss/generate.py --config manga-colorist/inference.toml

For your own page or a folder of pages, override the reference input and output directory:

uv run --no-sync python scripts/moss/generate.py \
  --config manga-colorist/inference.toml \
  --ref_image /path/to/your/page.png \
  --out_dir output/my-colorized-pages

The commands above use Bash line continuation. In PowerShell, put the command on one line. The generator keeps the reference aspect ratio and caps the output area at approximately one megapixel. The INT8 transformer used for these examples is handled by the upstream generator's model downloader.

Settings that matter

Setting Value used in the examples
Base model OpenMOSS-Team/MOSS-Image-Base
Prompt comic
Prompt format text
Prompt tail auto, resolving to the checkpoint's sure setting
Text encoder BF16
LoRA strength 1.0
Steps / CFG / seed 30 / 3 / 7
Sampler / shift ER-SDE / 5
Negative-branch skipped layer 22
Transformer / residual stream INT8 / FP32
Reference scale / aspect 1 / target
Reference timestep / isolation 0 / enabled
Reference position / vision conditioning aligned / disabled
Output area cap 1024 x 1024, preserving page aspect ratio
Acceleration Compiled blocks and Spectrum enabled

Keep the fixed prompt, text format, BF16 text encoder, and checkpoint prompt-tail behavior when reproducing these examples. The reference image is part of the trained conditioning path. A generic text-to-image LoRA loader or an ordinary img2img pipeline does not reproduce that path by itself. See the upstream MOSS image-to-image documentation for integration details.

Training

The adapter was trained on 2,134 black-and-white / synthetically colorized manga-page pairs from 42 series. The color targets use a cinematic anime palette, and every pair uses the caption comic.

Item Value
Adapter LoRA, rank 32, alpha 128
Adapted projections q, k, v, o, feed-forward up/gate and down projections
Optimizer / learning rate Muon / 5e-5
Total training 3 epochs, 6,402 updates
Schedule 1 epoch, then 2 additional epochs initialized from that checkpoint
Resume behavior Weights resumed; optimizer state restarted
Training transformer / LoRA weights INT8 stream / BF16
Caption / text encoder comic / BF16
Saved adapter EMA weights, window 0.1
Hardware NVIDIA RTX 4090, 24 GB
File size 148,721,952 bytes (about 141.8 MiB)

The safetensors metadata records 4,268 steps for the resumed stage. The cumulative total, including the initial 2,134-step stage, is 6,402. Full release metadata is in training_summary.json.

Evaluation scope and limitations

  • None of the six displayed pages appears in the training manifest. Gintama, Solanin, and JoJo's Bizarre Adventure are also unseen at the series level. Other pages from A Silent Voice, Maison Ikkoku, and Hellsing were used in training.
  • These six examples are qualitative demonstrations, not a standardized benchmark or evidence of performance on every manga style.
  • The adapter can redraw faces, clothing, architecture, shading, and small details. It can substantially change the original artist's style.
  • Dialogue and panel layouts are often recognizable in these examples, but exact lettering and line preservation are not guaranteed.
  • Colors are inferred from the input and learned palette. They may differ from canonical character or costume colors.
  • The Solanin and Maison Ikkoku base-model prompts were corrected after inspecting the original pages, then rendered again. Their corrected-prompt results are shown. None of the LoRA samples was regenerated.

Per-example dimensions, prompts, seeds, and hashes are included alongside the PNGs. See the comparison methodology for additional details.

License and acknowledgements

The LoRA weights and original repository materials are released under Apache License 2.0. The base model is MOSS-Image Base by OpenMOSS, also released under Apache 2.0. Training and inference use sorryhyun/anima_lora.

Example manga artwork retains the rights of its respective creators and publishers. The adapter license does not relicense that artwork. Creator credits are listed in NOTICE.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Examples
Examples
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for moonrailgun/MOSS-MangaColorist-v1

Adapter
(67)
this model