8 mins read

Playground V2, An Amazing Diffusion Model Competing with SDXL

Trained from a brand new dataset, a recently released model seems to do better than SDXL. Playground v2 is a valid 1024px-native base diffusion model that offers itself as a good alternative to Stable Diffusion XL: it has the same architecture as SDXL, but it has been trained on a different dataset of images with a native resolution of 1024px.

Let’s take a look at who’s behind the development of this model and how to run it locally or online. I will also include a simple comparison with SDXL for some prompts and samplers.

A Preference Study

Playground is currently working on an image editor that integrates advanced AI research with practical product design. The aim is to provide users with the ability to generate professional-grade graphics, regardless of their level of expertise.

One of the results of the research at Playground has been a newly trained text-to-image model, currently named Playground v2. It can be used for commercial purposes, and as stated in some benchmarks provided by the research team, the model generates better images in terms of aesthetics compared to SDXL.

Study user reported by Playground

According to the details in the study report, users were given instructions to score image pairs (SDXL and Playground V2) based on both their aesthetic preference and the image-text alignment. Playground V2 was preferred 2.5 times more than the SDXL ones. This obviously doesn’t mean that it is automatically better, but it certainly shows that it has the potential to surpass SDXL, depending on the type of AI images you want to generate.

Playground V2 and SDXL

The study caught my attention, so I decided to compare Playground V2 with Stable Diffusion XL myself; an alternative way to generate AI images is always welcome. I will first do a comparison using ComfyUI, then I will also show how to run the model in InvokeAI, as well as directly online.

ComfyUI

Running Playground V2 in ComfyUI is extremely easy; you can use the default workflow with minimal components.

Before that, you also need to download the model from playgroundai on Hugging Face. There are two safetensor files (playground-v2.fp16.safetensors and playground-v2.safetensors): the f16 one is smaller in size so you can start with that one.

Then load the default workflow or any SDXL-compatible workflow, and select from the Load Checkpoint node the file that you just downloaded. Be sure that the files are located in ComfyUI\models\checkpoints\.

comfyui playground v2 ai

One important thing to remember is that Playground V2 likes a very low CFG scale in the KSampler node; values below 3 are good, otherwise, you may get oversaturated or ugly results.

Starting with a simple prompt and no other modifiers or LORAs, I ran a quick experiment comparing Playground V2 and SDXL using the same workflow. The prompt was: “portrait of a blonde man, looking at the camera, bokeh, f1.4.” If you are looking for inspiration, have a look at this article about SDXL prompts.

For the SDXL experiments, I had to increase the CFG scale again to values around 7-8, which normally work better with stable diffusion models.

playgroundv2 samplers test

The results are rather different, and I’m not sure which one I like more. Playground V2 seems more detailed but also too plastic-like or saturated, whereas SDXL is more delicate and natural, but some areas seem too smooth.

Trying a more generic prompt, I got these results. I set the steps at 25 and used Euler A. The prompt are a varion of <subject> amazing depth, masterwork, surreal, intricately detailed, bokeh, perfect balanced, artistic photorealism, smooth, great masterwork, beautiful, standing in grass, sunset.

Again, different results, but I would say comparable in quality. Maybe Playground V2 seems to be slightly better looking, but it really depends on the style that you want to achieve.

I don’t notice any difference in terms of speed with my GPU (3060 with 12GB of VRAM), so it’s only a matter of aesthetics and style: around 16 seconds for a 1024x1024px image.

Normally, SDXL is used in combination with a refiner, so I tried to run the same prompt using two different workflows: one being the default one for Playground V2, and the other using a refiner for SDXL.

Prompt: anime style, a grassy hill with cherry trees, during a red sunset, blue sky, japan

The refiner gives really good results, a noticeable improvement. Moreover, these experiments didn’t use any enhancements like LORAs, custom SD models, upscaling, etc., so what you see here can definitely be improved for both models.

Use it Online

If you want to try Playground V2 without setting up a local installation, you can try it for free on playground.com after registering with a Google account. You can choose to create on a canvas (useful for outpainting) or a more standard board.

I don’t know how many generations are possible for free. I tried just a couple of examples, so it’s worth a try! On the site, it is possible to choose different models or add optimizations, so feel free to experiment and make comparisons with SDXL.

InvokeAI

You can also run this model in InvokeAI, another popular UI that came out around the same time as the stable-diffusion-webui. It works well on Mac, and it provides an easy-to-follow installation through the terminal.

Go to the InvokeAI git repository and scroll down the page to find the installation section. There are several options: the easiest one is to simply download and unzip the zip file at the end of this release page (always check for the latest release!), and then run the InvokeAI-Installer\install file. The terminal will open and it will ask you several questions about your type of GPU and the installation preferences.

The first time, it will take a while to download the suggested models, but when the process is completed, just head to your installation folder and run invoke.bat (or the equivalent for other operating systems). Then the UI should open in your browser. If not, just paste the local address that you see in the terminal.

Now we need to import the Playground v2 model. We can either use the previously downloaded safetensor or import it by pasting the Hugging Face repository in the Model Manager > Import Models panel: playgroundai/playground-v2-1024px-aesthetic.

It will take a while to download more than 10GBs! If you prefer to use the playground_v2.safetensor, paste the full local path to your model file instead and choose Add model. It will convert it into diffusers format, and then you can select it in the UI.

invokeai playground v2

When everything is done, go back to the text-to-image section of the UI, and you are almost ready to generate. Be sure to select the right model in the left panel and reduce the CFG scale to 3 or less. Then you are all set; just hit Invoke, and the image will generate soon.

Conclusion

So now you know that there is another model that you can use to generate amazing AI images, with results comparable to SDXL. I think that this diffusion model is a valid alternative to SDXL, especially in some domains like vivid and intense compositions. Moreover, the team behind it seems to be motivated to improve their model, and hopefully, it will remain accessible for local generations even in the future.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.