Article also available on my personal site: https://ntcai.xyz/articles/applying_stable_diffusion_15_models_loras_controlnet_to_sdxl_with_comfyui_and_img2img/
Problem: I want to reuse SD 1.5 checkpoints and loras with SDXL to get 1024x1024px images.
Then use the img2img outputs to fine tune SDXL loras. (wip)
Solution:
https://ntcai.xyz/comfyui/sd_1-5_to_sdxl_1-0.json
This shows how to load and use img2img on sdxl output with a 1.5 checkpoint and optionally lora(s). Uses comfyui, based on https://github.com/SytanSD/Sytan-SDXL-ComfyUI .
Before(sdxl+refiner):
After(sdxl+ aniverse sd15 checkpoint + trip lora)::
Note: These outputs can be used to finetune SDXL
Detailed instructions
Install Comfy UI and Comfy UI manager https://github.com/ltdrdata/ComfyUI-Manager
Download the SD XL to SD 1.5 comfy JSON and import it sd_1-5_to_sdxl_1-0.json
Click “Manager” in comfyUI, then ‘Install missing custom nodes’
Restart ComfyUI
Install SDXL (directory: models/checkpoints)
Install a custom SD 1.5 model (directory: models/checkpoints) https://civit.ai
Install your loras (directory: models/loras)
Restart ComfyUI
Select your checkpoints
Queue!
FAQ
Why?
To translate all the cool work from sd1.5 to sdxl.
My img2img result is distorted
The lora strength in my example is set high, try to set it lower.