The simplest way to self-host Playground v2.5 Aesthetic. Launch a dedicated cloud GPU server running Lab Station OS to download and serve the model using any compatible app or framework.
Direct Download
Download model weights for local inference. Must be used with a compatible app, notebook, or codebase. May run slowly, or not work at all, depending on your system resources, particularly GPU(s) and available VRAM.
Forge is a platform built on top of Stable Diffusion WebUI to make development easier, optimize resource management, speed up inference, and study experimental features.
Automatic1111's legendary web UI for Stable Diffusion, the most comprehensive and full-featured AI image generation application in existence.
Model Report
playgroundai / Playground v2.5 Aesthetic
Playground v2.5 Aesthetic is a text-to-image model built on SDXL architecture, using Energy-Based Diffusion (EDM) for improved color and contrast. It features balanced aspect ratio handling (9:16 to 16:9) and human-guided training for enhanced facial details and textures. Benchmarks show FID score of 4.48 vs SDXL's 9.55.
Explore the Future of AI
Your server, your data, under your control
Playground v2.5 Aesthetic, released in February 2024, represents a significant advancement in text-to-image generation, building upon the Stable Diffusion XL (SDXL) architecture while introducing novel methods to enhance aesthetic quality. The model utilizes two fixed, pre-trained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L) and is available under the Playground v2.5 Community License.
Key Improvements and Features
The developers focused on three primary areas of enhancement:
Color and Contrast: Using the Energy-Based Diffusion Model (EDM) framework, Playground v2.5 achieves superior color vibrancy and contrast compared to its predecessor's Offset Noise method. This improvement is particularly noticeable in generating pure-colored images and placing subjects on solid backgrounds.
Multi-aspect Ratio Generation: Through a refined data pipeline ensuring balanced sampling across different aspect ratios, the model excels at generating high-quality images across various dimensions, avoiding the biases seen in models like SDXL.
Human Preference Alignment: Drawing inspiration from Emu, the model implements a novel approach focusing on facial details, eye features, hair texture, and overall lighting, resulting in significantly fewer visual errors in human depictions.
Performance and Benchmarks
User studies and benchmarks demonstrate Playground v2.5's superior performance compared to both open-source and closed-source competitors. The model outperforms SDXL by a factor of 4.8x and PixArt-α by 2.4x in aesthetic quality, while also surpassing DALL-E 3 and Midjourney 5.2.
On the MJHQ-30K benchmark, Playground v2.5 achieves significantly lower FID scores (4.48) compared to Playground v2 (7.07) and SDXL (9.55), with particularly strong performance in the "people" and "fashion" categories.
Technical Implementation and Usage
The model supports generation at 1024x1024 resolution and various aspect ratios. It employs two different schedulers:
EDMDPMSolverMultistepScheduler (default) with a recommended guidance_scale of 3.0 for sharper details
EDMEulerScheduler with a recommended guidance_scale of 5.0 as an alternative option
