Playground v2.5 – 1024px Aesthetic Model - AI Image Models Tool

Overview

Playground v2.5 – 1024px Aesthetic Model is a diffusion-based text-to-image model optimized for high aesthetic quality at 1024×1024 resolution and for non-square aspect ratios. According to the model card and technical report, it is a latent diffusion model that follows an SDXL-like architecture and conditions on two fixed text encoders (OpenCLIP-ViT/G and CLIP-ViT/L). The authors highlight three development focuses: improved color/contrast via noise-schedule changes, balanced multi-aspect-ratio training, and alignment with human perceptual preferences (technical report published Feb 27, 2024). ([huggingface.co](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic?utm_source=openai)) The model is distributed openly with a Playground v2.5 Community License and is available for direct use via Hugging Face Diffusers and an API endpoint on Replicate. The Hugging Face pipeline notes EDMDPMSolverMultistepScheduler (EDM-formulation of DPM++ 2M Karras) as the default for crisper detail (recommended guidance_scale≈3.0), and an EDMEulerScheduler option (guidance_scale≈5.0) is also supported. Community and platform feedback praises its aesthetic output and FID improvements, while some users have reported img2img/init-image color/tint artifacts and occasional platform availability changes. ([huggingface.co](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic/blob/main/README.md?utm_source=openai))

Key Features

• High-aesthetic 1024×1024 outputs with portrait and landscape aspect ratio support
• Latent Diffusion architecture closely following Stable Diffusion XL patterns
• Dual fixed text encoders: OpenCLIP-ViT/G and CLIP-ViT/L for robust prompt conditioning
• Recommended EDMDPMSolverMultistepScheduler and EDMEulerScheduler support in Diffusers
• Open-source weights, Hugging Face Diffusers integration, and Replicate API endpoint
• Demonstrated FID and user-study gains vs SDXL, DALL·E 3, and Midjourney 5.2 (paper & model card)

Example Usage

Example (python):

from diffusers import DiffusionPipeline
import torch

# Install diffusers>=0.27.0 (or the main branch) and dependencies before running.
pipe = DiffusionPipeline.from_pretrained(
    "playgroundai/playground-v2.5-1024px-aesthetic",
    torch_dtype=torch.float16,
    use_safetensors=True,
    variant="fp16",
)
pipe.to("cuda")

prompt = "A cinematic portrait of a woman, dramatic lighting, ultra-detailed"
# EDMDPMSolverMultistepScheduler is recommended for crisper detail; guidance_scale≈3.0 is a suggested default.
image = pipe(prompt, num_inference_steps=28, guidance_scale=3.0).images[0]
image.save("playground_v2_5_out.png")

Pricing

Replicate lists an approximate runtime cost of about $0.16 per run (varies by inputs and hardware). The model weights are open-source so you can run locally to avoid API costs. Source: Replicate model page.

Benchmarks

MJHQ-30K overall FID (playground-v2.5): 4.48 (Source: https://replicate.com/playgroundai/playground-v2.5-1024px-aesthetic)

MJHQ-30K overall FID (playground-v2): 7.07 (Source: https://replicate.com/playgroundai/playground-v2-1024px-aesthetic)

MJHQ-30K overall FID (SDXL-1-0-refiner): 9.55 (Source: https://replicate.com/playgroundai/playground-v2.5-1024px-aesthetic)

Replicate runtime (typical prediction time): ≈113 seconds on NVIDIA A100 (80GB) (varies by input) (Source: https://replicate.com/playgroundai/playground-v2.5-1024px-aesthetic)

Replicate approximate cost per run: ≈$0.16 per run on Replicate (varies by inputs) (Source: https://replicate.com/playgroundai/playground-v2.5-1024px-aesthetic)

Key Information

• Category: Image Models
• Type: AI Image Models Tool