What models or files do I need, and where do they go?

This workflow relies on custom Qwen image layer decomposition nodes. Install the node repo as instructed by its README and place the required model weights in the paths it specifies (typically under ComfyUI/models or a subfolder the node creates). Then restart ComfyUI so the nodes detect the weights.

It feels slow—how can I speed it up?

Reduce Steps (e.g., from 50 to 30–40), lower the working size to 640 px in ImageScaleToMaxDimension, and keep CFG around 3.5–4.0. These changes noticeably cut runtime with modest quality tradeoffs. For the final pass, switch back to 1024 px and a few more steps.

Do I have to write a prompt?

No—the PrimitiveStringMultiline prompt is optional. However, short, concrete descriptions of the elements you want (e.g., "character, cape, sword, sky") often produce more accurate and neatly separated layers than running with defaults.

How do I get crisper edges and fewer halos?

Use a larger input (1024 px), increase Steps slightly (e.g., +10), and avoid very high CFG values. After decomposition, use the refinement/selection node to clean transitions. If needed, lightly feather or contract the mask in your compositor for final polish.

Qwen Image Layer Decomp for Illustration v1

Qwen Image Layer Decomp for Illustration v1 is a guided layer-separation workflow for turning a flat illustration into editable layers and masks. It routes a loaded image through ImageScaleToMaxDimension to a recommended working size (640 px for speed, 1024 px for detail), then drives a custom Qwen-based decomposition stage that predicts per-element mattes you can save as PNGs. A PrimitiveStringMultiline node provides an optional, high-level text prompt so you can nudge the model toward the objects you want isolated (for example: "character, foreground props, midground buildings, sky").

Technically, the core is implemented with two custom nodes (f754a936-daaf-4b6e-9658-41fdc54d301d and 1ea4fa01-a81b-4420-a03c-f60019e69a92). The first performs the Qwen image layer decomposition using diffusion-style sampling controls (the author’s baseline is 50 steps, CFG 4.0), trading runtime for mask quality. The second is used for selection/refinement of the predicted layers (cleaning edges or exporting specific passes). LoadImage and SaveImage handle I/O, while MarkdownNote contains on-canvas guidance and links. The result is a practical, prompt-guided layer extraction pipeline that’s ideal for compositing, recoloring, and 2.5D motion work.

API

Use this workflow from code

Every Comfy workflow is a JSON graph. The payload below is this workflow, exactly as ComfyUI runs it — fetch it from the URL, keep it in version control, or load it in ComfyUI and run it node by node.

a39b591d8c3b.json

Fetching workflow JSON…

Run it from TypeScript or Python with the Comfy SDK. The same code targets Comfy Cloud or a ComfyUI you host yourself — only the base URL changes.

// Install (beta)
npm i @comfyorg/sdk

// Run this workflow (TypeScript)
import { Comfy } from "@comfyorg/sdk";

const client = new Comfy({ apiKey: "comfyui-..." });
const wf = await client.workflows.fromFile("workflow_api.json");
const job = await client.run(wf);
await job.getOutputs("<output-node-id>")[0].toFile("output.png");

The SDK takes a workflow in API format: open this workflow in ComfyUI and use File → Export Workflow (API).

Comfy Cloud API access requires a plan with an API key.

SDK docs Get an API key

FAQ

Frequently Asked Questions