Stability AI released Stable Diffusion 3 Medium
Finally you Can Download Stable Diffusion 3 Here
At the beginning of this month, during the Computex Taipei 2024 event, Stability AI made an exciting announcement regarding their latest model. Today, they have officially provided the download method for Stable Diffusion 3 Medium. This new release is eagerly anticipated by many in the AI and tech communities.
To access Stable Diffusion 3 Medium, users can now visit the HuggingFace website where detailed instructions and necessary files are available for download. This development marks a significant milestone for Stability AI as they continue to advance the capabilities of their diffusion models.
Let go to Stablility AI’s Huggingface page:
https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main
First of all you need to agree using this model in non-commerical only.
Now you can download Stable Diffusion 3 Medium checkpoint model Now!
What Different of these Model?
sd3_medium.safetensorsincludes the MMDiT and VAE weights but does not include any text encoders, you need to downloadT5XXLtext encoder yourself.sd3_medium_incl_clips_t5xxlfp8.safetensorscontains all necessary weights, including fp8 version of theT5XXLtext encoder, offering a balance between quality and resource requirements.sd3_medium_incl_clips.safetensorsincludes all necessary weights except for theT5XXLtext encoder. It requires minimal resources, but the model's performance will differ without theT5XXLtext encoder.T5XXLis a new text encoder from Google, it has emerged as a powerful technique in natural language processing (NLP).
Official Workflow in ComfyUI
Remember update ComfyUI to latest version before start.
You need to download sd3_medium.safetensors and put it into the /models/checkpoints .
You also need to download clip_g.safetensors, clip_l.safetensors and t5xxl_fp16.safetensors from text_encoders folder and put them into the /models/clip .
Work with normal workflow without loading text encoder
If you do you also can use normal workflow by using sd3_medium_incl_clips.safetensors, it already included text encoder.
If you want to use T5XXL in normal workflow, you can try sd3_medium_incl_clips_t5xxlfp8.safetensors, it already included T5XXL fp8 .
When using the T5XXL model, the ability to understand natural language prompts is significantly enhanced. However, this improvement comes at the cost of requiring more memory. Since Stable Diffusion 3 Medium interprets prompts differently compared to models like SDXL, it will likely take more time to learn how to fully exploit the potential of Stable Diffusion 3.
The distinct way in which Stable Diffusion 3 Medium processes prompts means that users will need to invest time in experimenting and understanding the nuances of this new model. This learning curve is necessary to achieve the best possible results and to leverage the advanced capabilities that Stable Diffusion 3 Medium offers.