Camping
As I wrote in Mountain Air, I have been experimenting with a process I called “Multipass”, where I combined multiple checkpoint models in a single render. You can do that by changing the checkpoint model in Automatic1111’s hires fix. But what’s even more powerful is to adapt the creativity of SDXL with SD 1.5 models.
There’s no easy way to do this in Automatic1111, so I made these in ComfyUI using my own workflow. I don’t really want to share my spaghetti workflow, so here’s what I did in text:
- Render SDXL image without Refiner. You can use Refiner but since I’m not using it as a final image, I didn’t want to waste time.
- VAE Encode image to Latent. You can’t use the SDXL latent directly, so re-encoding the image into Latent.
- Denoise image using SD 1.5 checkpoint. Denoising with 0.5 works well. Adjust this higher / lower as needed.
Everything in this set was done this way. Specifically, I used these settings:
Pass 1
- SDXL Base, 30 steps, Euler a
- SDXL VAE
Pass 2
- Airfuck’s Brute Mix, 30 steps, DPM++ 2M SDE, 0.5 Denoise
- vae-ft-mse-840000-ema-pruned
Upscale
- 4x_NMKD-Siax_200k
FaceDetailer
- bbox/face_yolov8m.pt
I applied face detailer after upscale, so there’s more resolution for eye detail.