corrupted image outputs with safetensor

#5
by Waldo00 - opened

Getting a lot of glitchy artifacts on latest a11, is this an a11 upstream issue or is this an issue with the safetensor / model?

realistic photo of a panda sitting in a forest, waterfall, rain, bamboo Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 312736957, Size: 1024x1024, Model hash: 0bf1ce6b06, Model: SSD-1B, VAE hash: 551eac7037, VAE: sdxl_vae_0.9.safetensors, Version: v1.6.0-263-g464fbcd9

image.png

realistic photo of a panda sitting in a forest, waterfall, rain, bamboo Steps: 20, Sampler: DPM++ 2M, CFG scale: 7, Seed: 255731624, Size: 1024x1024, Model hash: 0bf1ce6b06, Model: SSD-1B, VAE hash: 551eac7037, VAE: sdxl_vae_0.9.safetensors, Version: v1.6.0-263-g464fbcd9

image.png

tried without VAE too, but that didn't change it, same metadata on SDXL looks fine

Segmind org

Likely a safetensors ckpt error. We'll look into it.

converting via: https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_sdxl.py worked, might be worth re-doing for official .safetensor

Segmind org

We've used that only for the conversion, Will check and update soon.

Segmind org

We'll be adding A1111 compatibility soon. Thanks for reporting!

Segmind org

ComfyUI support has been added.

Segmind org

Automatic1111 support has been added, please use this checkpoint and tell if the artifacts still persist.

Fixed, nearly identical to my pruned version too, good job

Waldo00 changed discussion status to closed

Sign up or log in to comment