Last week I got an email about the release of Stable Diffusion 3.5. As you might remember, the release of Stable Diffusion 3 was a big flop. To be honest, I wasn’t sure if we would se another AI model from Stability AI and the community has mostly switched over to running Flux from Black Forest Labs.
But now Stable Diffusion 3.5 has been released, and I’ve spent some time trying it out. Two new models have been released so far, and a third one is expected in a few days. Here is what’s new with SD 3,5.
Stable Diffusion 3.5 Base Models
Stable Diffusion 3.5 Large (download)
8 billion parameters, with superior quality and prompt adherence, this base model is the most powerful in the Stable Diffusion family. This model is ideal for professional use cases at 1 megapixel resolution.
Stable Diffusion 3.5 Large Turbo (download)
A distilled version of Stable Diffusion 3.5 Large generates high-quality images with exceptional prompt adherence in just 4 steps, making it considerably faster than Stable Diffusion 3.5 Large.
Stable Diffusion 3.5 Medium (to be released on October 29th)
2.5 billion parameters, with improved MMDiT-X architecture and training methods, this model is designed to run “out of the box” on consumer hardware, striking a balance between quality and ease of customization. It is capable of generating images ranging between 0.25 and 2 megapixel resolution.
Difference From Stable Diffusion 3
- Stable Diffusion 3.5 is under the community license, which is not as strict as the SD3 license
- Several base models (at least 3 to begin with)
- Community has been given the ability to create fine-tuned models
- Community has been given the ability to train LoRa for SD 3,5
- Optimized performance (especially SD 3,5 medium and Turbo)
- Creates people of different skin tones without the need to prompt for it
- Retain ownership of the media generated without restrictive licensing implications
- Even though you can create detailed NSFW images with the base models of Stable Diffusion 3,5, it’s not censored the way SD 3 was. And since it’s open for the community I’m sure there will be fine-tuned models and Loras that enables fully NSFW images, for those interested.
Visual Comparison
Since it is generative AI’s we’re talking about, I have also made some visual comparison between SD 3,5, Flux, SDXL, Pixart, Playground and Sana.
Images
Flux
Stable Diffusion
Flux
Pixart, Playground & SDXL
Stable Diffusion 3.5
Flux
Stable Diffusion 3.5
Conclusion
My personal opinion so far is that Stable Diffusion 3.5 is a huge improvement from SD 3, but I need to test the models some more. Especially I have to test them with Loras and other enhancements. Right now as they are, I don’t believe that SD 3.5 is as good as Flux when it comes to quality. I still see a lot of disfigured hands and fingers in images generated with SD 3.5, and even though it sometimes happens with Flux as well it’s not nearly as frequent. Flux on the other hand has the issue that without using Lora, most people look that same when it comes to facial features and structures.
I do believe that Stable Diffusion 3.5 will be on par with Flux, or at least close to, once the community has gotten some time to develop Loras and fine-tuned models. It’s just a matter of time.
Pretty much all women have the Flux chin, unless you use Lora to get around it.
Don’t forget to visit my Gallery. I try to update it on a regular basis.