What this workflow does

👉 This is not a fast image up-scaling process, the goal is to add detail at upscale for text to image generation.

How to use this workflow

👉 Select your checkpoint(s), vae(s), input text and queue prompt.

Key points.

1. Iterative Mixing Node - https://github.com/ttulttul/ComfyUI-Iterative-Mixer, adds noise to add (context?) specific details.

2. Iterative Upscale - https://github.com/ltdrdata/ComfyUI-Impact-Pack, easy multiple step ksampling.

3. Dynamic Thresholding to use a large CFG for initial image generation.

4. Kohya Deep Shrink to use a larger resolution for initial image generation (this is also used throughout the up-scaling processes.

5. FreeU_V2, arguable in its benefits but can improve the image generation focus which can be perceived as beneficial.

6. Self-Attention Guidance, similar benefits to FreeU_V2.

7. Fast Groups muting for identifying a seed for up-scaling.

8. The KSamplers are all defaulted to "fixed" so no need to rerun the entire workflow if adjustments are needed before the upscale sampler.

9. The Resolution avg node is used to determine the downscale factor for each PatchModelAddDownscale node (reroutes aren't working for math nodes hence the mess with the splines).

Tips about this workflow

👉 -I've had best results with using an ancestral sampler and a higher than typical number of steps in the denoising of the up-scaled latent-PixelKSampleUpscalerProvider.

-Self-Attention Guidance doesn't play nice with atypical resolution sizes, if you encounter an error make sure both width and height are divisible by 64 or disable the Self-Attention Guidance node.

-To use with SDXL checkpoints, change the two checkpoints, change the 2 vae used, change the Resolution avg input to 1024, bypass or remove the tile controlnet in the upscale section, and change the NNLatentUpscale type to SDXL in the upscale section.

🎥 Video demo link (optional)

👉 [Please add here]