Manga Colorizer — Qwen-Image-2.1 LoRA (early WIP)

A LoRA for Qwen-Image-2.1 (edit model) that colorizes black-and-white manga pages. Give it a B&W page — it returns a fully colored version. Trained on ~1,600 real manga pairs: officially colored One Piece / Bleach / Naruto chapters, JoJo's Bizarre Adventure Parts 5–7 (Color-ban + matching B&W scans), and a set of colored doujinshi pages.

This is an early, rough version (v0.1). A rework of the dataset and training is already planned — expect noticeable improvements in future updates. Please read the known issues below before downloading.

How to use

Works with any Qwen-Image-2.1 edit workflow (ComfyUI template "Image Edit — Qwen Image 2.1" works as-is).

Basic colorization (no reference):

• image_1: your B&W manga page

• prompt: color <image1> manga panel

• LoRA strength: 1.0 for maximum coverage, 0.7–0.85 for more natural, conservative colors (recommended — see issues below)

• cfg 1, euler/simple, 25–30 steps (the official Qwen pipeline uses 40–50; 6-step turbo LoRAs work but produce grainier results)

Reference mode (character consistency):

• image_1: B&W page, image_2: any colored page featuring the characters

• prompt: color <image1> manga panel. <image2> is color reference

• The model pulls palette from the reference — useful for batch-colorizing a chapter so characters stay consistent without per-page prompts.

The model was trained with this exact prompt format, so keep the <image1> / <image2> placeholders intact.

Known issues (current version)

• Speech bubble text gets mangled. The model redraws text instead of preserving it, and generates gibberish lettering. If text matters, plan to composite the original B&W lettering back over the result.

• Colors can be chaotic and oversaturated. The model "guesses" palettes, so it may assign arbitrary colors (red hair, golden armor, green cape) with no source to justify them. Lower LoRA strength (0.6–0.8) reduces this at the cost of weaker coverage.

• It can invent content on empty areas — plain white backgrounds may get filled with scenery or fake titles/logos.

• Small details drift between panels — eye color and small accessories may change within the same page.

• Trained at 640×480, so fine detail and lettering are its weakest points.

Roadmap

Planned rework: higher training resolution, dataset cleanup (filtering oversaturated targets, adding "keep empty background" examples), fewer epochs with checkpoint selection focused on text preservation and palette stability.

Notes

• LoRA rank 32, trained with DiffSynth-Studio, 3 epochs / ~4.9k steps.

• The dataset contains adult (NSFW) doujinshi colorizations as part of the training data.

• All manga artwork belongs to their respective authors/publishers; this is a research/fan project.