Eric Qwen Encode (Native)

The Decompose node hides all of Qwen's internals behind a diffusers pipeline. The pack's native branch doesn't - it exposes the pieces, and that means you need a node that turns text plus a reference image into real conditioning for the Qwen-Image family. That's EricQwenEncode: a Qwen-Image-Edit-style text encoder with layer-oriented defaults, built for ComfyUI's native graph.

What it does

You feed it a clip (a Qwen CLIP model - the text encoder that comes with a Qwen-Image checkpoint), plus positive_prompt and negative_prompt. The default positive is literally "decompose this image into separate layers with transparent backgrounds" - the pack assumes you're here to make layers, though you can obviously change it.

The interesting part is the optional image input. If you connect a reference image, it gets scaled to about 1MP, passed through as a vision input to the text encoder, and - if you also connect a vae - encoded to a latent that's attached to the positive conditioning as reference_latents. That's the Qwen-Image-Edit mechanism: the reference image steers the edit, the text tells it what to change. The outputs are standard positive and negative CONDITIONING, ready for a KSampler.

Functionally it's a cousin of ComfyUI's built-in TextEncodeQwenImageEdit, but with the image path optional and the defaults pointed at decomposition rather than generic editing.

Where it sits in the native workflow

The native layered chain goes: reference image → Add Alpha → Encode → Multi-Layer Latent → KSampler → Latent Cut to Batch → VAE Decode → Layer Extract → Save. Encode is the brain; the Multi-Layer Latent node sets up the frame structure the model samples.

Two honest warnings. First, this branch is experimental - the pack itself says so, and the native Qwen-Image-Layered path only got ComfyUI support a day after the model dropped in December 2025, so you're on the newer end of the compatibility curve. Second, this node doesn't load anything for you: you need a Qwen CLIP wired in, and if you want the reference-image path you also need the RGBA VAE from the pack's loader. If the image input stays empty, you just get plain text conditioning - which may be fine for you, but know that's what's happening.

Install: ComfyUI Manager (search "Eric Qwen") or git clone https://github.com/EricRollei/Qwen_Layers_Diffuser_Pipeline_Comfyui into custom_nodes, then restart.