Layer Diffuse Decode (RGBA)

This is the decode node from the pack's simplest example workflow, and the one most people should start with. It does exactly what the plain LayeredDiffusionDecode does - reconstruct the alpha channel that the patched sampler hid in the latent - but instead of handing you RGB and a separate mask, it returns a single IMAGE with the alpha already embedded. Save that with a PNG Save node and you have a transparent cutout, full stop.

How it works

Same transparent VAE decoder as the sibling nodes, same two required inputs, so the wiring is identical:

KSampler ─→ VAEDecode ─┐
          └─ (samples) ├→ LayeredDiffusionDecodeRGBA → PNG Save

Feed it the KSampler's latent as samples and the VAE-decoded RGB as images. It decodes the alpha from the latent, inverts it (alpha = 1 − mask, since the raw decoder emits a mask where white is transparent), and joins the channels back into one RGBA image via ComfyUI's JoinImageWithAlpha. No mask wrangling, no second output to keep in sync.

Inputs and outputs

• samples (LATENT) - the finished latent from the KSampler running on your LayeredDiffusionApply-patched model.
• images (IMAGE) - the ordinary VAEDecode result. Both inputs must be present; this node won't decode from the latent alone.
• sd_version - SDXL or SD15, matching your checkpoint. Controls which transparent-decoder weight loads (vae_transparent_decoder.safetensors for SDXL, layer_sd15_vae_transparent_decoder.safetensors for SD 1.5). Both auto-download to ComfyUI/models/layer_model.
• sub_batch_size - default 16, decode chunk size. Lower it if you're VRAM-crunched.

Output: a single IMAGE in RGBA.

When to reach for it vs. the plain decode

If you're making transparent PNGs for compositing or product shots, take this one - it's the least moving parts. Reach for LayeredDiffusionDecode instead when you want the alpha as a real MASK object (to blur, feather, threshold, or run through mask-based nodes), or when a downstream tool mangles RGBA and you'd rather keep the channels separate yourself.

Same traps as the rest of the family: latent dimensions must be multiples of 64 (the decoder asserts it - a 1000×1000 generation will fail here), most ComfyUI nodes silently flatten RGBA if you route through them, and the whole pack is SD 1.5/SDXL-only with no movement since roughly early 2025. Budget an extra 2–3GB of VRAM over a plain generation. But when it works, the edges are clean by construction - no background-removal network guessing where the hair ends.