ComfyUI with a 3 pass workflow. No lora/inpaint/detailer/manual edit. Batch gen for 1st pass at low steps as seed img noise inject sd 3.0 to 5.0 (sometimes different chkpt is used), then 2 step upscaled both w/ and w/o model. Prompts are gibberish or none. Other things i tweak alot are upscale method, different prompt/scheduler/sampler/steps(denoise)/seed/cnet for each pass. FreeU 1.2/0.9/1.0/0.2