Qwen Image Layer Decomposition

Turn a source image into a batch of generated layer images, guided by an overall image description. Upload your image and replace the example description “guitar” with a description of its objects, appearance and arrangement, including relevant partly occluded elements. The description guides the whole decomposition; there are no individual layer prompts or controls assigning objects to particular layers. For a first run, keep the maximum input dimension at 640 pixels, request two layers and use one decomposition per run. Leave the negative prompt empty unless you want to discourage particular attributes. Keep the sampling defaults: 20 steps, CFG 2.5, Euler sampler, simple scheduler, denoise 1 and sampling shift 1. The maximum input dimension sets the reference image’s longest dimension in pixels; the other dimension follows its aspect ratio. Output dimensions derive from this resized reference rather than independent width and height controls. Resize method changes the interpolation applied before reference encoding. More layers, larger images or more decompositions per run increase processing demands. Multiple decompositions use the same source and description, and their generated layers are returned together as an image batch. Use seed -1 for randomization on each run, or a fixed seed to compare settings with the same sampling noise. CFG controls text-conditioning influence. Sampler and scheduler select the denoising algorithm and noise schedule; sampling shift separately adjusts the model sampling schedule. Denoise controls sampling from an empty layered latent, so lowering it is not a source-image preservation strength. Results are generated image layers rather than an editable layered document. Sweet Tea returns the output batch through its image preview transport; this workflow does not permanently save images through ComfyUI SaveImage. The source loader’s RGB image output supplies the reference; uploaded alpha masks are not connected or preserved as input masks. The model, text encoder and layered VAE remain fixed to the source configuration.

Turn a source image into a batch of generated layer images, guided by an overall image description. Upload your image and replace the example description “guitar” with a description of its objects, appearance and arrangement, including relevant partly occluded elements. The description guides the whole decomposition; there are no individual layer prompts or controls assigning objects to particular layers. For a first run, keep the maximum input dimension at 640 pixels, request two layers and use one decomposition per run. Leave the negative prompt empty unless you want to discourage particular attributes. Keep the sampling defaults: 20 steps, CFG 2.5, Euler sampler, simple scheduler, denoise 1 and sampling shift 1. The maximum input dimension sets the reference image’s longest dimension in pixels; the other dimension follows its aspect ratio. Output dimensions derive from this resized reference rather than independent width and height controls. Resize method changes the interpolation applied before reference encoding. More layers, larger images or more decompositions per run increase processing demands. Multiple decompositions use the same source and description, and their generated layers are returned together as an image batch. Use seed -1 for randomization on each run, or a fixed seed to compare settings with the same sampling noise. CFG controls text-conditioning influence. Sampler and scheduler select the denoising algorithm and noise schedule; sampling shift separately adjusts the model sampling schedule. Denoise controls sampling from an empty layered latent, so lowering it is not a source-image preservation strength. Results are generated image layers rather than an editable layered document. Sweet Tea returns the output batch through its image preview transport; this workflow does not permanently save images through ComfyUI SaveImage. The source loader’s RGB image output supplies the reference; uploaded alpha masks are not connected or preserved as input masks. The model, text encoder and layered VAE remain fixed to the source configuration.