It's a ComfyUI workflow.This is my attempt at porting the SD 1.5 old idea to Z-image turbo. The idea is to use a model as a refiner, take image, add a bit of noise, diffuse and repeat a couple of times. It's a kind of gradual refinement... This way you can change overall image style while keeping the subject and composition somewhat intact. Like, for example, take your plasticky Illustrious render and make it a bit more real. Z-Image Turbo seems a perfect fit as it's fast and supports high resolution output.So far it's just an experiment. I'm not claiming it's successful, but looks promising to me.There are 6 instances of Ksampler in serial, with convenient on/off switches. rgthree nodes need to be installed. Mind the upscaler, bypass if not necessary.On sample images: input image is on the left, result on the right...A note: it looks like the model likes big prompts and can handle a lot of input. Use elaborated prompt to enforce desired style. You can use Qwen LLM to expand the prompts, it's good at it and also the TE for this model is based on it, so they should understand each other well. I guess. You can also use Qwen to create subject prompt from image, just feed it the input image and ask to make prompt out of it, then add a separate paragraph for the style.

Pixeldust Z turbo refinery

Added power lora loader and image post-processing node.

Files with SHA256: 1a9406d1d8e09bd69ddbd13410abeff9347b2be9f65dd31c9e22aeadf2360a0a