Description
This is a massive ComfyUI workflow that helps to improve and simplify detailing and upscaling of stable diffusion images by enabling easy and automatic masking along with prompt refinement for multiple characters at once, allowing users to do highly selective inpainting/detailing/upscaling without need for any manual masking, and minimal prompting.
Features:
- Automatic segmentation and masking for SFW content using natural language with SAM3.
- Automatic detection and bounding of NSFW anatomy using YOLO models.
- Supports detailing up to four characters simultaneously.
- Supports up to 3 LORAs per character/detailer pass combination.
- Universal LORA and Upscaler LORA.
- Centralized control of key functions across all characters and/or body parts: Sampler, Steps, CFG, Denoise Levels and more.
- Automatic prompt enhancement based on characteristics of the subject. (Generated with Gemma4)
- Previews for every step of refinement, allowing for easy pinpointing if something goes wrong.
- Up to 6 refinement passes per character: full body, face and hair, clothing, breasts, genitals, and feet.
Uses
The workflow primarily enhances the fine details and overall quality of stable diffusion images - it is also possible to use it for face swap and clothing swap workflows.
Bias, Risks, and Limitations
Currently, the prompt enhancement logic is pretty cisnormative and heteronormative. This may be improved on in a future release by adding more subject characteristic flags, but the prompt enhancment logic is already kinda complicated. For the time being, for generating trans or gender nonconforming characters, consider disabling the prompt enhancement block and prompt for those manually.
How to Get Started with the Model
Simply drag and drop the JSON file into ComfyUI! Read the notes in the workflow for an explanation of what everything does.
Before Example:
After Example:
Credits:
Big thanks to catlover1937 over at CivitAI. This workflow is a massive expansion and adaptation of the "Face Detailer" and "SD Upscale" sections of their excellent (and much more reasonable in scale) ZIB + ZIT Simple Workflow. This workflow would not have come into existence without that essential DNA to work from.