Playground v2.5 - 1024px Aesthetic Model

Introduction

Playground v2.5 is an advanced diffusion-based text-to-image generative model designed to produce highly aesthetic images. It generates images at a resolution of 1024x1024 and supports various aspect ratios. It is an improvement over its predecessor, Playground v2, and claims to surpass other state-of-the-art models like SDXL, PixArt-α, DALL-E 3, and Midjourney 5.2.

Architecture

The model is built on a latent diffusion architecture, utilizing two pre-trained text encoders: OpenCLIP-ViT/G and CLIP-ViT/L, similar to the architecture of Stable Diffusion XL. The default scheduler used is the EDMDPMSolverMultistepScheduler, with a guidance scale of 3.0, which is suitable for generating crisper fine details.

Training

Playground v2.5 was developed by Playground and underwent extensive user studies to enhance its aesthetic quality, multi-aspect ratio generation, and human preference alignment, particularly for people-related images. The model demonstrates superior performance in both qualitative and quantitative benchmarks, such as the MJHQ-30K, indicating its alignment with human aesthetic preferences.

Guide: Running Locally

1. Install the necessary packages:

   pip install diffusers>=0.27.0 transformers accelerate safetensors
   

2. Run the model:

   from diffusers import DiffusionPipeline
   import torch
   
   pipe = DiffusionPipeline.from_pretrained(
       "playgroundai/playground-v2.5-1024px-aesthetic",
       torch_dtype=torch.float16,
       variant="fp16",
   ).to("cuda")
   
   prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
   image = pipe(prompt=prompt, num_inference_steps=50, guidance_scale=3).images[0]
   

3. Cloud GPUs: For optimal performance, consider using cloud-based GPUs from providers like AWS, GCP, or Azure.

License

The model is released under the Playground v2.5 Community License. Details of the license can be found here.