← sources

I trained a LoRA that makes MOSS-Image output real transparent PNGs (proof of concept, 39 training images)

source: reddit (reddit:StableDiffusion)published 2026-10-24 06:37ingested 2026-10-24 10:59

retained text

It's late, so Claude wrote this for me, but basically: there's a project, [https://huggingface.co/trmz/flux2-klein-alpha](https://huggingface.co/trmz/flux2-klein-alpha), where they trained Flux Klein to work with RGBA. Since the VAE is the same, I wondered whether MOSS-Image could produce RGBA too, and surprisingly, it can! It only takes a small dataset and a short training run. Anyway, here's the text Claude wrote:



Hi! I've been experimenting with getting MOSS-Image to generate images with a real alpha channel, and the proof of concept turned out better than I expected, so I'm sharing it.



How it works: you end your prompt with "isolated on a transparent background" and decode with the RGBA VAE from trmz/flux2-klein-alpha. You get a transparent PNG straight from the sampler: no background removal, no chroma key. It even keeps partial transparency on glass, fairy wings, gems and the Eiffel Tower's lattice.



Why a LoRA is needed: the RGBA VAE can already decode MOSS latents (same FLUX.2 latent space), but plain MOSS latents give garbage alpha. The LoRA teaches MOSS to produce latents that carry the alpha signal.



It was trained on only 39 images in \~40 minutes on a single 16 GB GPU, so it's not perfect. You'll sometimes see faint halos or a barely-visible box where the background alpha isn't exactly zero. But it generalizes well: none of the characters in the examples were in the training set.



Everything is on Hugging Face: the LoRA, a small ComfyUI node to load the 4-channel VAE, a workflow, 36 RGBA examples, and research notes on how I built the dataset (Klein + Extract-4B and Qwen Image 2.1 as alpha "teachers").



[https://huggingface.co/paralaif/MOSS-Image-Alpha-LoRA](https://huggingface.co/paralaif/MOSS-Image-Alpha-LoRA)





document fetched 2026-10-25 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xe1gse/i_trained_a_lora_that_makes_mossimage_output_real