QwenImage Layered V2 Decompose (compat)

Like the loader's (compat) twin, this is the same code as QwenImage Layered V2 Decompose registered under a legacy name. The pack's NODE_CLASS_MAPPINGS maps both names to the same class, so a workflow saved before the V2 rename still loads it without breakage. If you see it in your graph, it is the real Decompose - same inputs, same outputs, same behavior.

So the short version of the article: it runs the DiffSynth Qwen-Image-Layered V2 pipeline on your input image and returns one isolated RGBA layer. Required inputs are pipeline (from the loader), image, prompt, seed, steps (default 10), cfg_scale (default 1.0), width/height (default 1024), layer_num (default 0) and use_input_size (default True); optional are negative_prompt, context_image, target_mask and remove_mask. Output is a single layer_image in RGBA. The prompting rule matters as much here as on the V2 node: describe the whole image, never write "cut out" or "isolate" - the brush decides the region, the prompt describes the scene.

Why two names at all?

The pack started as a local experiment, the author renamed the nodes to make the "V2" pipeline explicit, and the compat aliases were kept so nobody's early workflows exploded on update. It's a courtesy, and it's the sort of thing that tends to disappear once the pack settles down. Which one should you use? If you're adding nodes by hand, pick the V2 names - they're the ones the README, the example workflows, and future fixes will target. If a saved workflow already has this node, just leave it alone.

Everything else you need to know

All the substance lives on the V2 page: how the context_image brush works, the red/green/yellow channel logic, the cfg_scale around 1.0 recipe, the soft-conditioning caveat (the brush steers but doesn't pin pixels), and the heavy setup - DiffSynth-Studio installed manually, official Qwen text encoder/tokenizer and layered VAE cached, allow_official_download defaulting to False. None of that changes based on which name the node answers to. This alias gets you to the same place with the same caveats; it just answers to an older name when it gets there.