Eric Qwen Layer Prompts

Prompting a model to "split this into layers" sounds easy until you try to get consistent results out of it. The wording of the positive and negative prompt decides whether you get clean, separated layers or a muddy mess with halo artifacts everywhere. EricQwenLayerPrompts exists so you don't have to learn that prompt engineering the hard way - it hands you a curated set of decomposition prompt pairs.

How it works

A preset dropdown picks the strategy, and the node returns a matching (positive, negative) pair as two STRING outputs. The presets cover the common jobs:

• layer_decomposition: the general-purpose one - background layer first, then foreground elements in depth order.
• foreground_background: clean subject-from-background separation, with negatives aimed at halo edges and incomplete cuts.
• subject_extraction: every distinct object becomes its own layer.
• text_extraction: pull text/typography out, negatives aimed at broken letters and blur.
• color_separation: group elements by dominant color.
• custom: you bring the prompts via custom_positive and custom_negative.

The num_layers input (2–10, default 4) gets substituted into the prompt templates, so the model actually knows how many layers you're asking for. Set it to match the num_layers on your Decompose or MultiLatent node and the prompt and the latent agree with each other.

Where it plugs in

Two places, depending on which half of the pack you're using. In the diffusers workflow, feed the positive output into the Decompose node's prompt input and negative into negative_prompt - though honestly the Decompose node's built-in auto-captioning already handles this when you leave prompt empty. Where this node really earns its keep is the native ComfyUI workflow: those prompts go into the Eric Qwen Encode node, which builds the actual conditioning for a Qwen-Image-Edit-style graph.

The honest caveat

These are starting points, not magic. The diffusers pipeline and the native Qwen-Image path respond differently to phrasing, so treat the presets as a good default and expect to tweak the custom slot for your specific images. Text extraction in particular is a where-the-rubber-meets-the-road job - you may want to bump num_layers up for complex posters.

Standard pack install: ComfyUI Manager (search "Eric Qwen") or git clone https://github.com/EricRollei/Qwen_Layers_Diffuser_Pipeline_Comfyui into custom_nodes, then restart. No model load - it just formats strings.