Qwen Image Layered

Qwen Image Layered is the oddball in the pack's WaveSpeed section, and it's genuinely cool: feed it a single image and it decomposes it into separate RGBA layers - foreground subject, background, objects - with transparency built in. Out the other side you get a batch of layer images and a matching batch of alpha masks, ready for compositing work. It's like asking a model to do the layer separation you'd normally do by hand in Photoshop.

It's a WaveSpeed-hosted Qwen capability, so the mechanics are a billed API call: the image goes up, the model reasons about the scene's depth and object boundaries, and the result comes back as multiple layers. You control the granularity with num_layers - 2 to 8, default 4. More layers means finer separation (and, realistically, more chances for the model to split something oddly); fewer means chunkier, safer cuts. The prompt input lets you steer the decomposition - describe what matters ("keep the person separate from the car, merge the road and sky") in Chinese or English.

The inputs and outputs

• image - the image to decompose, as a URL or path (a string, not a tensor)
• num_layers - 2 to 8, default 4
• prompt - optional guidance for how to split the scene
• seed - cache control; randomize for a fresh decomposition each run
• enable_sync_mode / enable_base64_output - completion and encoding behavior
• client - optional WaveSpeed client

The two outputs are the point: images (IMAGE batch, the layer RGBs) and masks (MASK batch, one alpha per layer). Wire them into compositing nodes - the mask batch is exactly what the pack's Region Mask-style tooling or any mask-based compositor wants. Layer 1 through N correspond positionally between the two outputs, so images[i] pairs with masks[i].

Install and setup

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Or "ERPK Custom Nodes" from ComfyUI Manager, restart, and set a WaveSpeed API key in ERPK Settings (right-click canvas).

Troubleshooting and honest expectations

"Image must be provided" is the validation error you'll hit if you queue it empty - no surprise there. The bigger caveat is expectations: model-based layer separation is impressive but not pixel-surgical. A busy scene with overlapping subjects will come back with imperfect edges, and num_layers doesn't guarantee the kinds of layers you want - you might get "person / background / shadow" instead of "person / car / road / sky." That's what the prompt guidance is for. If your source is a clean product shot or a single subject on a simple background, this node can save you serious masking time. For chaotic scenes, budget for cleanup. The transparency genuinely is real though - that's the part people don't expect, and it opens up the whole compositing toolbox.