Stable Diffusion 3.5 Medium: Features and Community Adoption

Stable Diffusion 3.5 Medium is poised for strong community adoption due to faster generation and improved image quality.

### Stable Diffusion 3.5 Medium: Community Adoption and Performance Insights
* 🤖 The SD 3.5 Medium model shows promise for adoption by the open-source community, potentially allowing it to **compete with Flux** due to its ability to **generate 1440x1440 images natively** and **faster inference speeds**.
* ⚙ The model has a token limit of around **500**, which is important to note as exceeding the **256 token limit** can lead to breakdowns in adherence and additional artifacts in the output.
* 📊 There are two downloadable models: one without text encoders at just over **5 GB**, and another provided by Comfy that includes text encoders at approximately **11.5 GB**.
* 🎨 In terms of output quality, the prompt adherence is strong regarding **color and contrast**, although improvements are still needed for **hand rendering**.
* 🖼 Compared to Flux, SD 3.5 Medium excels in **photorealism and skin textures**, although text generation remains **less coherent than the large model**, requiring **multiple generations for satisfactory results**.
* 🛠 The introduction of the SLG workflow with a **skip layer guidance node** may help improve **anatomical accuracy**, particularly in problematic areas like **legs**, although it is not a complete solution.
* 💻 Performance tests on a Ryzen 5800 X show varying inference times, with **1440x1440 images taking around 1 minute and 36 seconds to generate**, highlighting the model's **efficiency despite its smaller size**.
* 🔗 Stability AI has successfully released the SD 3.5 Medium model, which is smaller at **2.5B parameters**, and offers **dual attention blocks** in the first 12 Transformer layers, enhancing **context length for prompt processing**.

so Stability AI stuck to their word and released Stable Diffusion 3.5. Medium, as always, there'll be a link to everything you need in the description below. And before we look at the workflow in some examples, there are some things I want to point out. This model is 2 and 1/2 B, so smaller file size generation times should be a lot faster as well. But what's also different is the native resolution is 1440. So theoretically, you can generate larger images right off the bat. Under the model description, they talk about how there's dual attention blocks in the first 12 Transformer layers. Keep this in mind for later. This is actually going to be very important in terms of context length. You can go about 256 tokens. If you don't know what tokens are, it's just a way where prompts are broken down. If I take a look at this prompt here, we see the token count is at 62. So the word astral has two tokens. Each of the colors represents one token. So wave is a token, half is a token, even

the commas take up these tokens. One of the limiting factors of SDXL was that you were limited to 77 of these tokens, so long prompts didn't work with SDXL. I believe Flux is at 500 roughly, and I could be wrong, but I think SD 3.5 large is around that 500 token mark as well. Now, that's important to know because if you go over that 256 FA limit, things might start to break down, adherence could suffer, and you can also get some additional artifacts. The other thing to be aware of, they say here that it's a different training data set than the large model, so your prompts won't necessarily be similar. Now, there are two models you can download. The first one is just the model itself, not including the text encoders, and it's just over 5 GB. The other model that's available was provided by Comfy, and this one includes the text encoders at about 11 1/2 GB. So, it's really up to you which one you want.

to run, you can put both of these models in your checkpoints folder. The workflows are pretty similar to what I showed you with SD 3.5 large, but I do want to point out there's another one called SLG example. This is the default SLG workflow, and they provide a before and after. What's different about this workflow is this additional node called skip layer guidance sd3. Make sure you're on the latest version of kyui. This was just implemented, I guess, today. So earlier, I mentioned about there being 12 layers. What this does, it skips seven, eight, and nine. And this node may help with anatomy deformities. It's not a cure for it, but it may help. I left it on its default setting, and I left all the default configurations together. This was the prompt that was provided. So here's the original image, and we see the legs are kind of messed up, but with this new node, we see that the legs are much better, right? Now, I'll be honest, I don't know how this all works. I just discovered it.

myself but try the default settings at first, and it looks like it's connected to the model sampling node, and then the output goes to model under your K sampler. Now the recommended settings are 30 to 40 steps, starting at CFG 4.5 and up, sampler DPM Plus+ 2 msgm uniform. I've tried Oiler beta and simple, and it works fine in terms of inference speed on my Ryzen 5800 X, 32 GB of RAM, 3060 TI, 8 GB of VRAM. For these times, I was doing 40 steps, 1280 by 720, 32 seconds, 896 X 1152, 38 seconds, 1024x 1024, 38 seconds, and 1440 by 1440, a minute and 36. These examples are just random prompts that I put together, comparing it between large and medium. As mentioned before, they're trained on different datasets, so you are going to get different results. This one was quite interesting because in the prompt, I had colored pencil, whereas large leans towards photo realism.

generally speaking, the prompt adherence is really good in terms of color and contrast. I have no complaints here, and it's no surprise. Again, with the hands, they could definitely use some improvements, although I would say on some of the characters I've done, the hands look pretty decent. But considering it's a 2 and 1/12b model and much smaller in file size, I could definitely see the community adopting this model, doing fine tunes and lauras, not to mention the quicker inference times. And just like the large model, the photo realism, the skin textures look better, I would say, than flux, whereas with flux, it tends to have that hyper-realistic plastic look. In terms of text, it performed pretty well with just a few words. When I tried doing the full-word stable diffusion 3.5 medium, it took quite a few generations just to get one decent one. So I would say the text coherency is far back from the large model. It definitely needs some improvement, but if you keep it to like

two to four words, it should be fine. So, I'm curious what you think about SD 3.5 medium. Do you think the open-source community is going to adopt this model to use for training and luras? Personally, I think this is how SD is going to gain some ground on flux. The fact that you can generate 1440 by 1440 images natively and the quicker inference speeds are definitely a plus. Now, in case you haven't seen the SD 3.5 large video that I did last week, make sure to check it out right here. Until next video, my friends, I'll see you when I see you.

Stable Diffusion 3.5 Medium Smaller & Faster!

Monzon Media

5 has been released by Stability AI, featuring a smaller file size and faster generation times.

The model's native resolution is 1440, allowing for larger image generation.

It also incorporates dual attention blocks in the first 12 Transformer layers, enabling a context length of about 256 tokens for prompt breakdown.

The model sampling node is connected to the output, with recommended settings of 30 to 40 steps, starting at CFG 4.

The Oiler beta and simple samplers work well for inference speed on a Ryzen 5800 X with 32 GB of RAM and a 3060 TI.

Different datasets for large and medium models result in varying outputs, with the large model leaning towards photo realism.

1/2B model shows good prompt adherence, with decent hand representation and better photo realism compared to the larger model.

Text coherency is lacking, requiring multiple generations for decent results.

The new SLG workflow, with an additional node for skip layer guidance, may help improve anatomy deformities.

5 medium's open-source model is expected to gain traction for training and luras, potentially competing with Flux.

The model offers native 1440 by 1440 image generation and quicker inference speeds.

It has a token limit similar to Flux, with two downloadable models available: one at just over 5 GB and the other, provided by Comfy, at 11.