LoRA trained on 39 images in 40 minutes makes MOSS-Image output real transparent PNGs
gatortux · reddit · 2026-10-24
A developer trained a LoRA that lets MOSS-Image emit PNGs with a genuine alpha channel straight from the sampler — no background removal or chroma keying needed.
- How: end the prompt with "isolated on a transparent background" and decode with the RGBA VAE from trmz/flux2-klein-alpha. That VAE can already decode MOSS latents (shared FLUX.2 latent space), but vanilla latents yield garbage alpha; the LoRA teaches MOSS to produce latents carrying the alpha signal.
- Cost: trained on just 39 images in 40 minutes on a single 16 GB GPU. Partial transparency survives on glass, fairy wings, gems, and the Eiffel Tower's lattice, and none of the example characters were in the training set.
- Caveats: occasional faint halos or a barely-visible box where background alpha isn't exactly zero.
- Resources: the LoRA, a ComfyUI node for the 4-channel VAE, a workflow, 36 RGBA examples, and dataset notes (using Klein + Extract-4B and Qwen Image 2.1 as alpha "teachers") are all on Hugging Face.
More from coding & agent
- Voice-controlled auto-advancing slides go viral; Tejas Kumar built the same in 2018 — TejasKumar_ · 2026-10-24
- Solo founder: after 6 months I stopped reviewing my AI agents' code — Al_Grigor · 2026-10-24
- Building an enterprise-wide custom agent platform: which UI/harness to use? — Electronic_Sort_6627 · 2026-10-24
- Dev wowed by Gemini 3.1 Sol: one prompt yields a glowing rainbow portal shader — Dimillian · 2026-10-24
- Hacker ran AI agents to breach 100+ sites at $50 each; researcher shares 5 Grok Bot defense prompts — alex_verem · 2026-10-24
- antirez now runs every side project with Claude and GPT cross-checking each other — antirez · 2026-10-24