Curiosity · AI Model

Stable Diffusion XL 1.0

Stable Diffusion XL 1.0 (SDXL) is Stability AI's 2.6B-parameter text-to-image latent diffusion model, released under CreativeML Open RAIL++-M in July 2023. It ships as a base model plus a refiner U-Net that denoises the final steps for extra detail, uses two CLIP text encoders for richer prompt conditioning, and supports native 1024×1024 generation. SDXL 1.0 became the de facto open image backbone, spawning thousands of LoRAs, fine-tunes (Juggernaut, RealVisXL), and ControlNets.

Model specs

Vendor: Stability AI
Family: Stable Diffusion
Released: 2023-07
Context window: 1 tokens
Modalities: text, vision

Strengths

Open weights under CreativeML Open RAIL++-M
Huge ecosystem of fine-tunes and LoRAs
Native 1024×1024 with refiner
Runs on consumer GPUs (≥12 GB VRAM)

Limitations

Weak text-in-image rendering versus newer DiT models
Human anatomy sometimes degrades at odd poses
Behind SD3 / Flux / Imagen 3 on absolute quality
RAIL licence has use restrictions

Use cases

Self-hosted image generation
LoRA / DreamBooth fine-tuning
ControlNet-driven composition
Open backbone for product avatars and creative tools

Benchmarks

Benchmark	Score	As of
Human preference vs SD 1.5	preferred ≈68% of the time	2023-07
PartiPrompts win-rate vs Midjourney v5.1	roughly on par at launch	2023-07

Frequently asked questions

What is SDXL 1.0?

SDXL 1.0 is Stability AI's 2.6B open-weights latent diffusion text-to-image model, released in July 2023 with a base+refiner architecture and native 1024×1024 output.

How is SDXL licensed?

Under the CreativeML Open RAIL++-M licence — free for most research and commercial uses, with RAIL-style use restrictions (no deceptive/illegal deployment).

Is SDXL still state of the art?

No — newer models like Stable Diffusion 3, Flux, and Imagen 3 beat it on absolute quality, but SDXL remains widely used because of its mature ecosystem.

Sources

SDXL 1.0 announcement — accessed 2026-04-20
SDXL paper (arXiv) — accessed 2026-04-20