Transparent PNGs straight out of MOSS-Image. End your prompt with isolated on a transparent background, decode with the RGBA VAE, and the sampler gives you an image with a real alpha channel: no chroma key, no background removal, no masking.
Trained with black-and-white manga in mind (characters, creatures, props, ink effects, sound effects, speech balloons), but it also works in color.
Everything you need (ComfyUI)
Files, raw RGBA examples, trailer and full notes: huggingface.co/paralaif/MOSS-Manga-Alpha-Image-LoRA
This LoRA →
models/loras/RGBA VAE
vae/flux2_klein_alpha_rgba_vae.safetensors(in the Hugging Face repo; originally trmz/flux2-klein-alpha, Apache-2.0) →models/vae/Load RGBA VAE (Flux2) custom node + minimal workflow: the attached zip (also on Hugging Face). Stock ComfyUI can't load a 4-channel VAE.
MOSS-Image SFT + Qwen3.5-9B text encoder: Comfy-Org/MOSS-Image
8-step DMD LoRA:
lora.safetensorsfrom lodestones/MOSS-Image-SFT-8-steps-dmd, renamedmoss_image_sft_8step_dmd.safetensors
Tips
Always end the prompt with
isolated on a transparent background, and describe one subject, not a place.Say "man" or "boy" explicitly when it matters; ambiguous roles sometimes come out as women.
Semi-transparent areas (smoke, plastic umbrellas, ink sprays) can carry a faint purple tint from the RGBA VAE.
The preview images here are flattened onto colored backgrounds because Civitai doesn't show transparency. The raw RGBA PNGs are on Hugging Face.
Training
MOSS-Image Base, ai-toolkit, 503 RGBA images, rank 32, AdamW 8-bit 2e-4, step 1500 of 3000, one RTX 5060 Ti 16 GB. The training images are synthetic: 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. The artist tags were removed from the captions, but the LoRA may still carry some of those styles.
License
Apache-2.0. Credits: MOSS-Image (OpenMOSS, Apache-2.0), MossAD-Demo (RicemanT, Apache-2.0), RGBA VAE (Xavier Jara / trmz, Apache-2.0, method AlphaVAE), 8-step DMD (lodestones, Apache-2.0), ai-toolkit (Ostris, MIT).
Also available for Krea 2: paralaif/Krea2-Manga-Alpha-Image-LoRA
Description
LoRA strength 1.0. MOSS-Image SFT + 8-step DMD LoRA: 8 steps, CFG 3.5, er_sde / simple, SkipLayerGuidance on single layer 22 (0–50 %), negative worst quality, bad anatomy, watermark, jpeg artifact. 832×1216, 1216×832 and 1024² tested. Also works on MOSS-Image Base and MossAD.
Decode with the 4-channel RGBA VAE (in the Hugging Face repo) through the Load RGBA VAE (Flux2) node from the attached zip.