Stable Cascade Prompt Following Is Amazing — This Model Has Huge Potential — High Resolutions Uses Lesser VRAM & Still Very Fast — Check Comments For

Stable Cascade Prompt Following Is Amazing — This Model Has Huge Potential — High Resolutions Uses Lesser VRAM & Still Very Fast — Check Comments For

You can download our scripts here : https://www.patreon.com/posts/98410661

Stability AI recently released Stable Cascade. You can read here : https://github.com/Stability-AI/StableCascade

I have developed an amazing 1 click installer and Gradio app for this newest amazing model.

This model can be considered as next level of Stable Diffusion.

The gradio APP I developed supports low VRAM and works great on even 8 GB GPUs

Saves every generated image automatically in outputs folder and many a lot of improvements

Kaggle not working right now due to FP16 bug and I have reported it to be fixed. Hopefully after that notebook will work great

Batch size 4, 1536x1280 resolution it / s is 1.7 on RTX 4090

Batch size 1, 1024x1024 resolution it / s is 12.14 (encoder) / 10.6 (decoder) on RTX 4090

So 1 image takes like 4 seconds on RTX 4090 for 1024x1024


Josef Kulovany

zCHG - WeCharg - StealthMachines - Love Letter Media - Josef K Media - trisandcats.com

1 年

“I been to the desert on a horse with no name…”

Jin W.

Generative AI Prompt Curator | Pharmacist | Connector of people across technology, healthcare and finance

1 年

This is amazing.

Johan Fredriksson

AI & Machine Learning Developer @ AFRY

1 年

Also changing "Decoder Guidance Scale (CFG)" above 1 seem to generate this error: Traceback (most recent call last): ?File "Q:\StableCascade\venv\lib\site-packages\gradio\queueing.py", line 495, in call_prediction ??output = await route_utils.call_process_api( ........................ ?File "Q:\StableCascade\venv\lib\site-packages\diffusers\pipelines\stable_cascade\modeling_stable_cascade_common.py", line 304, in forward ??x = x + self.effnet_mapper( RuntimeError: The size of tensor a (2) must match the size of tensor b (4) at non-singleton dimension 0 Also values 0-1 does no diffrence for the generation

回复
Johan Fredriksson

AI & Machine Learning Developer @ AFRY

1 年

Changing "Prior Inference Steps" does nothing. It still runs 29 prior inference steps and there is no difference in speed or the look of the image. Some code missing? :)

回复

要查看或添加评论,请登录

Furkan G?zükara的更多文章

社区洞察

其他会员也浏览了