SDXL Turbo generation + two step interactive selection upscale using preview chooser
created 10 months ago
img2img
upscale
38 nodes

7.8k

1.9k

Outputs
example image output 0 for workflow.
example image output 1 for workflow.
example image output optimized_comfy_workflows_user_uploads/fa6988df-c5e9-4d9c-852b-5db7af8f5130/assets/astro_00025_.webp for workflow.
Description

This is an example of how you can use SDXL Turbo to generate a lot of low-res images, and then upscale the the best options. Images are generated at the 512 resolution supported by SDXL Turbo and then are upscaled using a regular SDXL model (Zavychroma). The first upscale goes to 1024 resolution, and I have two different styles of upscale which you can pick whichever seems to be better. One is the iterative upscale, and the other uses an upscale model and then runs this though an image2image resample. After you pick one of the two it goes into a tiling 2x upscale using Ultimate SD Upscale.

Using preview chooser there are two choice points: after the initial image batch, and then after the two 1024 upscale options. This is nice because you can decide to give up and just reroll the dice without going through all the upscale, but you can also reroll at the intermediate choices, or change the upscale parameters, and the try again using the chooser restart.

WARNING: preview chooser will often dump mangled workflows into the images. I think this happens when you use the restart option. It seems to work consistently to save a JSON workflow.

Built-in nodes
Custom nodes
CheckpointLoaderSimple
FromBasicPipe
PixelKSampleUpscalerProviderPipe
VAEEncode
ImageUpscaleWithModel
ImpactKSamplerBasicPipe
KSamplerSelect
CLIPTextEncode
PreviewImage
VAEDecode
SDTurboScheduler
SamplerCustom
IterativeImageUpscale
ImageBatch
ToBasicPipe
UpscaleModelLoader
Get Image Size
PrimitiveNode
UltimateSDUpscale
EmptyLatentImage
SaveImage
Preview Chooser
Models
2
0
2
0