paralaif/MOSS-Manga-Alpha-Image-LoRA

MOSS Manga Alpha Image LoRA

Transparent PNGs straight out of MOSS-Image. Prompt as usual, end with , isolated on a transparent background, decode with the RGBA VAE included here, and the sampler gives you an image with a real alpha channel: no chroma key, no background-removal model, no masking.

It was trained with black-and-white manga in mind (characters, creatures, props, ink effects, sound effects, speech balloons), but it also works in color.

Trailer made only from outputs of this LoRA (MOSS-Image SFT), composited on colored backgrounds.

Gallery

30 raw RGBA outputs (MOSS-Image SFT, 8 steps, one seed each, nothing retouched). The page background shows through the transparent areas. Full-resolution files are in examples/.

What's in this repo

File	Put it in	Notes
moss_manga_alpha_image_lora.safetensors	models/loras/	the LoRA (rank 32)
vae/flux2_klein_alpha_rgba_vae.safetensors	models/vae/	4-channel FLUX.2 RGBA VAE from trmz/flux2-klein-alpha, redistributed unmodified (Apache-2.0)
comfyui/ComfyUI-RGBA-VAE/	ComfyUI/custom_nodes/	Load RGBA VAE (Flux2) node; the stock VAE loader cannot load a 4-channel VAE
comfyui/moss-manga-alpha-image-t2i.json	drag into ComfyUI	minimal workflow, tested with the files below

Not included (download from the original repos):

File	Put it in	Source
moss_image_sft_int8_convrot.safetensors	models/diffusion_models/	Comfy-Org/MOSS-Image
qwen3.5_9b_int8_convrot.safetensors	models/text_encoders/	Comfy-Org/MOSS-Image
lora.safetensors → rename to moss_image_sft_8step_dmd.safetensors	models/loras/	lodestones/MOSS-Image-SFT-8-steps-dmd

How to use (ComfyUI)

1. Copy comfyui/ComfyUI-RGBA-VAE into ComfyUI/custom_nodes/ and restart ComfyUI.
2. Place the files from the tables above and load comfyui/moss-manga-alpha-image-t2i.json.
3. Write your prompt and end it with , isolated on a transparent background. The captions followed black-and-white manga illustration of <subject>, isolated on a transparent background.
4. The saved PNG is RGBA.

Settings used for every image on this page

Model	MOSS-Image SFT int8 + 8-step DMD LoRA
LoRA strength	1.0
Sampler	8 steps, er_sde / simple, CFG 3.5, negative worst quality, bad anatomy, watermark, jpeg artifact
Guidance	SkipLayerGuidanceDiTSimple, single layer 22, 0–50 %
Resolution	832×1216, 1216×832, 1024²

It also works on MOSS-Image Base and MossAD: we ran the same 30 prompts on all three and every image came out with valid alpha. SFT gave the heaviest ink and the most seinen look, MossAD the lightest and softest.

Training

Base	MOSS-Image Base (int8 convrot), ai-toolkit with RGBA targets encoded by the RGBA VAE
Data	503 RGBA images, captions black-and-white manga illustration of <subject>, isolated on a transparent background
LoRA	rank 32, alpha 32
Optimizer / LR	AdamW 8-bit, 2e-4, shift timesteps, 1024 px, batch 1
Steps	3000 run, step 1500 released
Hardware	1× RTX 5060 Ti 16 GB

About the data. The training images are synthetic. They were generated with MossAD-Demo using its built-in artist-style tags, on a flat indigo background, cut out with a chroma key and reviewed by hand (373 rejected). The artist tags were removed from the captions, but the LoRA may still carry some of those styles.

Limitations

• Faint purple tint in semi-transparent areas (smoke, plastic umbrellas, ink sprays). The FLUX.2 RGBA VAE encodes transparency as a magenta shift, and some of it leaks into the color.
• Gender drift on ambiguous prompts: "office worker" or "devil hunter" sometimes came out as women; say "man" or "boy" explicitly.
• Very dense black subjects and full scenes are harder; describe one subject, not a background.

Credits and licenses

• MOSS-Image by OpenMOSS: Apache-2.0.
• MossAD-Demo by RicemanT (training-data generator): Apache-2.0.
• RGBA VAE by Xavier Jara, trmz/flux2-klein-alpha (method: AlphaVAE): Apache-2.0.
• 8-step DMD LoRA by lodestones: Apache-2.0.
• ai-toolkit by Ostris: MIT.
• This LoRA, the custom node and the workflow: Apache-2.0. See NOTICE.

See also the earlier proof of concept, paralaif/MOSS-Image-Alpha-LoRA, and the Krea 2 version, paralaif/Krea2-Manga-Alpha-Image-LoRA.