Playground v2.5 – 1024px Aesthetic Model

This is diffusion text-to-image model designed to generate high-aesthetic images of 1024x1024 pixels, including portraits and landscapes. It is the successor to Playground v2 and demonstrates superiority over models like SDXL, PixArt-α, DALL-E 3, and Midjourney 5.2 in user studies focused on aesthetic quality. According to user research, the model outperforms both open-source and proprietary alternatives in aesthetic quality and in the following categories:

• Support for multiple aspect ratios.
• Alignment with human preferences (especially in images involving people).
• In tests on the MJHQ-30K benchmark, the model achieved the lowest FID score (4.48, compared to 7.07 for its predecessor and 9.55 for SDXL).

Architecture: Built on Stable Diffusion XL, it uses two pre-trained text encoders: OpenCLIP-ViT/G and CLIP-ViT/L.

Announce Date: 16.02.2024
Parameters: 3.5B
Context: 77
VRAM requirements: 1.9 GB using 4 bits quantization, 3.8 GB using 8 bits quantization, 7.6 GB using 16 bits quantization
Developer: Playground
License: Playground v2dot5 Community

Use our pre-built public endpoints for free to test inference and explore Playground v2.5 – 1024px Aesthetic Model capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.