Description
Sausage.Redmond is here!
Introducing Sausage.Redmond, the first (i think) big finetune of wuerstchen made by the community!
I'm grateful for the GPU time from Redmond.AI that allowed me to make this LORA! If you need GPU, then you need the great services from Redmond.AI.
The model is not perfect but its really cool to gen!
I really hope you like this and use it.
To use it:
vladmandic/automatic: SD.Next: Advanced Implementation of Stable Diffusion (github.com)
or
warp-ai/wuerstchen-prior script.
You just need to change the
prior_pipeline = WuerstchenPriorPipeline.from_pretrained(
"warp-ai/wuerstchen-prior", torch_dtype=dtype
).to(device)To:
prior_pipeline = WuerstchenPriorPipeline.from_pretrained(
"artificialguybr/SausageRedmond-wuerstchen", torch_dtype=dtype
).to(device)If you like the model and think it's worth it, you can make a donation to my Patreon or Ko-fi.
Patreon:
https://www.patreon.com/user?u=81570187
Ko-fi:https://ko-fi.com/artificialguybr
BuyMeACoffe:https://www.buymeacoffee.com/jvkape
Follow me in my twitter to know before all about new models:
FAQ
Comments (13)
absolutely GREAT work, you are the first AI finetuner and maker who made a Würstchen model finetune, and it looks crazy good!! finally somebody that saw that the model had potential. finally some model that can start generating the actual image in very low image size to improove the speed and reduce the computing power that generates actually good images!!! but i have one question: is it blocked on image resolutions between 1024x1024 & 1536x1536 or does it generates other resolutions?
What's a 'Würstchen' finetune? Is it a different base model I'm not aware of? Will it work with XL LoRA, embeddings, etc?
@Polygon Hi! Würstchen is an open source AI project basically developed by Warp ai that uses a low resolution diffusion and upscaling algorithm that makes the model fast and power efficient. You can check everything at https://huggingface.co/warp-ai/wuerstchen
@Grilgo Thank you for sharing! Is it compatible with XL LoRAs & embeddings?
Can you run it on colab?
@Polygon no, it's not. It's not SDXL.
@Polygon Since they are two different models made by two different "companies" I don’t think so. Maybe warp ai and stability ai will collaborate one day and XL Loras will be compatible with Würstchen but as of now I don’t think so…
@taniasanders You can run base (non-finetuned) Würstchen on colab by forking this colab made by camenduru : https://colab.research.google.com/github/camenduru/Wuerstchen-colab/blob/main/Wuerstchen_colab.ipynb
If the crator or someone else doesn’t release the sausage redmond version on colab, no
Is this XL? Could you upload a safetensor?
Würstchen is an entirely different model that provides much better speed compared to SDXL.
Is it works with LORas?
Nice to see someone investing the time for a new architecture! I wonder if there is a way of interpolating between SD1.5 and Wuerstchen just like the current stream with SDXL https://github.com/city96/SD-Latent-Interposer https://github.com/dome272/Wuerstchen
P.S. discovered this through other repos https://github.com/Nerogar/OneTrainer
Latent interposer gives bad results actually. I see more promising to use compositing in ComfyUI.
1. Generate a background with one model (SDXL, SD1.5 or any other)
2. Generate a main subject on simple background with similar lighting.
3. Generate a mask with Masquerade Nodes > MaskByText
4. Use compositing node to combine BG and subject
5. Then use Img to Img with KSampler or UltimateUpscale to finalize.
Upscaling latents with any scaling alg gives ugly artifacts very similar to interposer.