Image-to-image LoRAs for MOSS-Image-Base, trained with anima_lora. The input image enters the DiT as a clean VAE reference latent. One folder per task.
| Folder | Task |
|---|---|
colorize/ |
colorize a B&W manga page or line-art illustration |
sanitize/ |
remove text, SFX, speech-bubble text and watermarks from an illustration |
swap_text/ |
re-letter a Japanese manga page in English or Korean |
decensor/ |
remove bar and mosaic censors |
| Task | Capabilities |
|---|---|
colorize |
Preserves text in manga, stable color assignment across scenes. line-art illustration colorization, captionable color; v5 takes per-character colour boxes. |
sanitize/ |
Detects and removes unnatural text, SFX, speech-bubble text and watermarks, preserving exact color tone. Removal targets vary by seed. |
swap_text/ |
English: natural rendering for both sfx and dialogue, Korean: readable dialogue, unstable sfx |
decensor/ |
blurry mosaic removal, incapable of free inpainting. |
Needs ComfyUI-MOSS-Toolkit 0.6.1 or
later, the moss-image node, and the
Spectrum KSampler pack for the
MOSS Spectrum KSampler. MOSS Caption Helper registers only when
anime-tools is installed in ComfyUI's environment
(pip install 'anime-tools @ git+https://github.com/sorryhyun/anime_tools').
colorize/moss_colorize_v5.safetensors in models/loras/.colorize/moss-i2i-colorize-v5.json and pick your page in Load Image.How the workflow is set up:
task colorize. The LoRA's metadata turns on
ref_isolate and ref_timestep 0; task colorize reads them and puts the page in at
full size.task colorize_v5 (Toolkit 0.13.0 or
later, anime-tools 0.9.4 or later) into a plain CLIP Text Encode, the page connected.
v5 copies the lettering from the page itself, so the caption carries no OCR text. For a manga
page, leave comic: the helper adds a box-only text item per line it reads, as in training.
For an illustration, type colour and series tags (touhou, purple eyes, purple hair, blue shirt); the helper keeps those and drops the rest.worst quality, greyscale, blurry, jpeg artifact.er_sde, uncond skip on block 22.On an older Toolkit, use task layout with comic or the tags. The lettering still comes
from the page.
Character colours on a manga page: write the layout JSON yourself in the helper's caption, a
character item per character with its box and colours. colorize_v5 keeps them and appends
the text boxes. Boxes are yxyx pixels with the origin at the centre of the output canvas.
Keep it on one line with no spaces, as the LoRA was trained on:
{"summary":{"caption":"comic"},"items":[{"label":"character","caption":"red hair, red sweater","bbox":[-150,40,400,520]}],"bbox_coord_space":"abs_centralized","bbox_format":"yxyx"}
With comic alone, a real B&W page gets little colour; character items add it.
v5 trained on control captions: colour tags for illustrations; for manga, comic, per-character
colour boxes, and box-only text items. It sees no OCR text. On held-out pages it keeps the
lettering at least as well as v4 and lands closer to the target colours. Trained on 3,334
colour / B&W pairs, half illustrations and half manga pages (rank 32, 10,000 steps, Muon + EMA).
v4 (moss_colorize_3k_v4.safetensors, workflow moss-i2i-colorize.json) is the OCR-caption
version: its helper runs task colorize with the page connected.
sanitize/moss_i2i_sanitize_v3.safetensors in models/loras/.sanitize/moss-i2i-sanitize.json and pick your image in Load Image.v3 leaves the rest of the image's tone alone: in the split above the two halves meet with no seam in the background, skin or hair, while the text on the right half is gone and the hearts (art, not text) stay.
How the workflow is set up:
task custom, ref_scale 1 (full-size reference). The
LoRA's metadata turns on ref_isolate and ref_timestep 0.task sanitize into a plain CLIP Text Encode.
It outputs no sfx, no dialogue, no watermark, the one caption the LoRA was trained on, so
don't type your own prompt.The output keeps the input's aspect ratio, with the area capped at resolution² (1024).
v3 was trained on 277 pairs, 39 of them clean → clean (rank 32, 2,216 steps, Muon + EMA). In the
second half of training a tone term on the untouched region holds the output's brightness
and colour to the input's; earlier versions could brighten and cool the whole image.
moss_i2i_sanitize.safetensors (v1, 49 pairs) and _v2 stay in the folder.
swap_text/moss_i2i_swap_text_v4.safetensors in models/loras/.swap_text/moss-i2i-swap_text-v4.json, pick the Japanese page in Load Image, and
write the prompt in MOSS Caption Helper.Needs ComfyUI-MOSS-Toolkit 0.12.0 or later (Caption Helper task swap_text) and anime-tools
in ComfyUI's environment.
How the workflow is set up:
task auto. The LoRA's metadata turns on ref_isolate
and ref_timestep 0; the page goes in at full size.task swap_text, the CLIP connected, its
conditioning output as the positive. v2 onward learned Korean with one token per Hangul
syllable, and the helper encodes the prompt the same way. The same prompt through a plain
CLIP Text Encode merges syllables into tokens the LoRA never saw (English is the same
either way).er_sde, uncond skip on block 22.The prompt is layout JSON with one text item per line to letter: the translated text and
its bbox in abs_centralized yxyx pixels of the output canvas (the page at its own aspect,
area capped at resolution²). The summary caption is comic, english text or
comic, korean text. A balloon with no item stays Japanese.
{"summary":{"caption":"comic, english text"},"items":[{"label":"text","caption":"text","text":"WHAT ARE YOU DOING?","bbox":[-600,250,-420,400]}],"bbox_coord_space":"abs_centralized","bbox_format":"yxyx"}
MOSS Caption Helper on task colorize with the page connected writes one item per Japanese
line it reads, boxed on the same canvas; replace each text with its translation and the
summary caption with one of the two above.
Each training pair is a Japanese page with the clean lettering of its translated edition pasted in. All versions: rank 32, modulation rank 16, Muon + EMA.
v4 (moss_i2i_swap_text_v4.safetensors, workflow moss-i2i-swap_text-v4.json): learned
line breaks. A \n in an item's text breaks the lettered line there; write the breaks the
balloon should have.
decensor/moss_i2i_decensor.safetensors in models/loras/.decensor/moss-i2i-decensor.json and pick your image in Load Image.The sanitize setup, with MOSS Caption Helper on task decensor: it outputs
no bar censor, no mosaic censor, uncensored, the LoRA's one training caption. Same
sampler (30 steps, CFG 4) and output size.
Trained on 48 censored / uncensored pairs (rank 32, 576 steps).
swap_text/moss_i2i_swap_text_v4.safetensors 149 MBswap_text/moss_i2i_swap_text_v3.safetensors 149 MBswap_text/moss_i2i_swap_text_v2.safetensors 149 MBswap_text/moss_i2i_swap_text.safetensors 149 MBcolorize/moss_colorize_v5.safetensors 149 MBdecensor/moss_i2i_decensor_v2.safetensors 149 MBsanitize/moss_i2i_sanitize_v3.safetensors 149 MBcolorize/moss_colorize_3k_v4.safetensors 149 MBcolorize/moss_colorize_3k_v3.safetensors 149 MBdecensor/moss_i2i_decensor.safetensors 149 MBsanitize/moss_i2i_sanitize.safetensors 149 MBsanitize/moss_i2i_sanitize_v2.safetensors 149 MBOpenMOSS-Team/MOSS-Image-Base