// the find
lllyasviel/sd-forge-layerdiffuse
[WIP] Layer Diffusion for WebUI (via Forge)
A Forge/A1111 extension that generates actual transparent PNG layers straight out of Stable Diffusion, instead of running background removal after the fact. It works with both SD1.5 and SDXL via LoRAs and a custom VAE encoder/decoder pair, and can also generate foreground+background+blend as a matched set. Aimed at people doing asset generation for compositing, game art, or design work who are already running Forge.
The core trick — a special VAE that decodes native alpha channel from latent space — actually produces clean edges on things like hair, fur, and glass that naive matting can't touch, and the demo images back that up. Model selection and download are automatic, so you're not hunting for the right checkpoint. The joint generation modes (fg+bg+blend in one batch) are a genuinely useful workflow for compositing since the layers stay coherent with each other rather than being generated independently and stitched.
Marked WIP with no versioned releases, and the README is really an accumulated pile of sanity-check screenshots rather than documentation — there's no concise reference for parameters or model files, you have to read through prompt examples to figure out what each mode does. The DPM++ sampler artifact issue is called out but unresolved in the README (workaround is 'use a different sampler'), which suggests rough edges in how the training distribution interacts with certain samplers. SDXL joint fg/bg generation is explicitly on hold due to 3x slower inference and 4x VRAM, so the more useful workflow is only fully available on SD1.5. Last push was August 2024 with no changelog since, so it's unclear if this is maintained or effectively frozen.