I decided to revisit my original rapid Turbo XL to SD1.5 workflow (https://civarchive.com/models/217808/cjs-super-simple-sdxl-turbo-to-sd-15-workflow) to harness the advancements that have been made over the last year for Stable Diffusion.

This workflow is no longer limited to Turbo only. You should be able to use any SDXL (or XL related) checkpoint with it. Similarly, if you're willing to tweak and test settings, you can keep this entirely limited to SDXL only, including tile upscale!

The speed here comes entirely from two Lora packs. PCM is the latest and greatest low-step lora set released as a sort of a successor to LCM. It is no longer limited to just the LCM sampler which offers a lot more flexibility. SPO stands for Step-aware Preference Optimization and is a sort of successor to DPO. The SPO locon lora is distilled from the SPO checkpoint for use on any SDXL checkpoint.

The workflow can be used for a lot of different things, but my favorite is live composition of images. If you wish to use this for live image composition, set seed to "New Fixed Random" and make sure the "Upscale" group is turned off/bypassed. In the side menu, check "Extra options", and then turn on "Auto Queue". Press "Queue Prompt" and type away. Once you've built an image you like, then you can turn off "Auto Queue" and activate the "Upscale" group.

From initial inference to final upscaled image at 2048x3072, total time for generation with my 4090FE is ~20 seconds.

Please note that there are limitations with this workflow. Namely, as with any Img2Img generation that using different models, your images will change some between each pass. If things aren't looking how you'd like them, try changing the seed or modifying your prompt. You can also look at adjusting the denoise level of the SD1.5 ksampler. I've generally found between 0.40 to 0.65 to work well. Anything higher or lower will result in blurry or chaotic images, so I don't recommend it. 0.60 denoise remains my preferred denoise level.

Hope you have fun! If there is demand for it, I may roll out a single SDXL checkpoint