Empty Qwen Image Layered Latent

In December 2025 Qwen released a model that had people posting "Photoshop will be useless in six months." That model was Qwen-Image-Layered, and its superpower is that it doesn't generate one flat image - it generates the image as separate RGBA layers, so you can edit a character without touching the background. This node creates the layered latent canvas that model samples into. It exists for exactly one workflow, and if you're looking at it, you're probably running that workflow.

Four inputs:

• width / height - default 640×640, step 16.
• layers - default 3. The number of foreground layers you want the model to separate. Each gets its own latent plane, so 3 layers means the output decomposes into 3 foreground elements plus a background.
• batch_size - default 1.

Output: a LATENT shaped [batch, 16, layers + 1, height/8, width/8] - 16 latent channels (the Qwen/Wan-family latent), 8× spatial compression, and a depth axis of layers + 1. That extra +1 is the background layer, always there. This is a genuinely different shape from every other empty latent in ComfyUI - the model's whole point is that this stack of planes is the layered image.

How it works

A normal image latent is one 16-channel plane. A layered latent is several planes stacked along a new dimension. Qwen-Image-Layered was trained to fill that stack so that each plane is a coherent RGBA layer - one plane holds the person, another holds the prop, the background plane holds everything else, and the alpha channel per plane is what lets you composite them in any order. The layers count is your say in how many foreground elements you're budgeting for: too few and the model lumps things together, too many and it invents layers to fill them. 3 is the sane default; the model can recurse and produce variable counts, but the empty latent fixes the initial budget.

Where it fits

The layered workflow: this node → KSampler (with the layered text encoder giving each layer a description) → VAE Decode → the stack comes out as per-layer images you can edit independently and re-composite. That's the "edit stays isolated to one layer" promise: change the character plane and the background plane never moved.

The honest reality check

• It only works with the Layered checkpoint. The regular Qwen-Image and Qwen-Image-Edit models don't understand the layers+1 shape - feeding this latent to them gives garbage or an error. If you don't have Qwen-Image-Layered loaded, this node is useless to you.
• It's a niche node with real downstream complexity. The decomposition is impressive, but using the layers afterwards (compositing them with correct alpha, in the right order) is where the workflow gets fiddly. The node is the easy part.
• The base Qwen VAE question follows you here. Qwen-Image's VAE (which Wan 2.1's VAE is compatible with) is the subject of the whole 2026 "over-smoothing" debate - if your layered output looks soft, that's the familiar Qwen VAE trade for text-legibility, not the layering.

One node, one model, one workflow - but it's the node that made "generate, then edit by layer" a real thing, and if you've seen the demo you know exactly why the Photoshop-obsolescence takes happened. Just make sure your checkpoint is the Layered one before you blame the node.