A ComfyUI workflow combining Qwen Image 2.1 text-to-image generation, an optional prompt enhancer, Ultimate SD Upscale, and three independently controlled detailers.
Built with a compact layout, intermediate previews, and explanatory notes to make adjustments easier.
Features
Optional prompt enhancement to expand your image description before generation.
Optional Ultimate SD Upscale with a dedicated ON/OFF switch.
Three detailers, applied in order: 01 Hands → 02 Feet → 03 Breasts.
SAM3.1 detection with masked crop refinement and stitching.
Separate controls for each detailer’s prompt, sampling settings, and denoise.
Bypass paths that preserve the incoming image when a detailer is disabled or no regions are detected.
Brown documentation nodes with tips, model downloads, and credits.
Processing order:
T2I generation → optional upscale → hands → feet → breasts → final preview.
Choose your own detailing model
The breast detailer uses Redcraft by AiMetatron in my default configuration, together with a Heretic text encoder and madebyollin’s texture-fix VAE.
Redcraft is simply my personal preference—not a requirement. You can use another suitable image model, including standard Qwen Image 2.1, by configuring the appropriate encoder, VAE, conditioning, and sampling settings. Different model architectures may require adapting the branch.
Why no face or eye detailers?
I intentionally excluded them after finding that additional refinement could alter facial features and unusual eye colors. The optional upscale still processes the whole image, so compare results with it disabled when preserving those details is especially important.
Getting started
The included README covers installation, required custom nodes, model downloads, controls, prompting tips, and troubleshooting. Downloads are separated into the default workflow configuration and standard Qwen alternatives.
All model and custom-node credits are included in the README and workflow notes. Model weights must be downloaded separately.
Start with the saved settings, inspect the previews, and adjust one stage at a time. Detailers can improve local detail, but correcting a difficult pose may still require refining the original generation.
Feel free to share your results and feedback!