download the YAML
use a webui that supports v-pred, turn on ZTSNR betas, and use CFG Rescale=0.7 (InvokeAI with diffusers link recommended because it handles this for you
use NON-K samplers (ZTSNR does not work with Karras)
if you don't YOU WILL GET BLACK/BROKEN/WEIRD IMAGES
Photography model, resumed from Pretty Girls Next Door with new training data - prompting tips can be found here. The model uses an OpenClip-H text encoder. The prompting style is halfway between SD1.5 and SDXL. Some of the training data is booru tags, some is moondream English sentences, some is hand-labelled.
Recommended settings
This is a v-pred model trained with ZTSNR. Use CFG Rescale = 0.7 and enable zero-terminal SNR betas. For best results, use InvokeAI.
Image size; 640x896
Steps; 50
Scheduler; DPM++ 3M or DPM++ 2M SDE (NOT KARRAS)
Description
FAQ
Comments (5)
I get moire, a black background with visible spots
are you using the YAML? also i forgot to write that it doesn't work with k-samplers
unfortunately I need a bit more detailed instructions to get this to work. I guess you can't use this with Automatic?
you should be able to. download the safetensors file, downlad the yaml and save it in the same folder, look for a setting in auto1111 called something like 'sd_noise_schedule' and set it to "Zero Terminal SNR".
it should work without the setting, you just need the yaml.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.