Sign In

Ultraspice

1.4k
20k
396
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
1,293
Reviews
Published
Apr 24, 2024
Base Model
SDXL Turbo
Hash
AutoV2
19AFC65557
This Stability AI Model is licensed under the Stability AI Non-Commercial Research Community License, Copyright (c) Stability AI Ltd. All Rights Reserved.

Please do share your creations if you make something that you like. We can all learn from each other.

For best results:

Start with 1024x1024 or 1024x768. Then put that image into controlnet and select Depth. Render again with the same (or different) prompt using double the resolution of the original (2048x1536). I recommend using the DPM++ SDE Karras sampler with a CFG scale between 1 and 2.5, and 5-11 steps. I've seen users make beautiful pics with other samplers, so it may be worthwhile to experiment to see what suits your desire.

VAE is included. Generally speaking, adding steps increases detail but if you want to use any sort of LoRA or hiresfix, a higher step count may fry your image. Lower cfg tends to follow the prompt better; higher cfg tends to give more contrast.

ADetailer and Highresfix may require slightly different settings than non-turbo models. I usually use a denoise strength around 0.2 but YMMV.

If you find a composition you like but the fingers are bad, save the seed and adjust the CFG scale by 0.1 or 0.01. They tend to successfully generate within a couple of tries.

1.6 improvements to contrast and composition.

1.5 improves color and detail.

1.4 merged in some of my favorite LoRAs.

1.2 Is more creative and tends to handle hands better.

Version 1.0 is functionally the same as 0.9t, the only notable difference being more life-like and saturated colors.

0.9t represents a huge leap over the previous iteration. I've merged in some of the realness of Juggernaut and the responsiveness of OpenDalleV1.1 and crammed it back down to Turbo(merge) speed.