← All models

Stable Diffusion XL 1.0

Image Available Full comparison ↗

stability-ai/sdxl · by Stability AI · Latent diffusion

Pricing — 1 offering(s)

Replicate

View on Replicate ↗

About this offering: Stability AI's SDXL 1.0 on Replicate, billed per second of GPU processing time rather than per image. It runs on an Nvidia L40S at $0.000975 per second, and a median image costs about $0.003 (Replicate model page, checked 2026-10-11).

A100 80GB (compute-billed)

No current price · last price $0.0014 / second, until 2026-10-10

  • $0.0032 / second Historical 2023-07-26 → 2023-08-15 · Nvidia A100 (80GB)
    Pre-reduction hardware rate. Start date approximated to the SDXL 1.0 release date; the blog documents this as…

    Pre-reduction hardware rate. Start date approximated to the SDXL 1.0 release date; the blog documents this as the rate before the cut.

  • $0.0014 / second Historical 2023-08-16 → 2026-10-10 · Nvidia A100 (80GB)
    Current public-model rate (~$5.04/hour). Replicate bills only active processing time; cold starts bill at the…

    Current public-model rate (~$5.04/hour). Replicate bills only active processing time; cold starts bill at the same per-second rate.

L40S (compute-billed)

  • $0.000975 / second Current 2026-10-11 → present · Nvidia L40S
    Replicate now runs SDXL on an Nvidia L40S at $0.000975 per second of processing time; a median run costs…

    Replicate now runs SDXL on an Nvidia L40S at $0.000975 per second of processing time; a median run costs about $0.003 (Replicate model page, checked 2026-10-11).

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Stable Diffusion XL 1.0 fits into a cost-aware routing setup

See how →

Capability profile

How Stable Diffusion XL 1.0 rates across core capability dimensions, with the task-level evidence behind each rating.

Artistic style range Strong

Vast community LoRA/fine-tune ecosystem effectively extends its range.

Photorealism Strong

Strong with the refiner and photoreal fine-tunes.

Prompt adherence Moderate

Below DALL·E 3 and FLUX on prompt following — see compositional-accuracy for the GenEval detail.

Compositional accuracy Moderate

GenEval (automated text-to-image alignment over six compositional task types — single object, two objects, counting, colours, relative position, attribute binding; 0–1 = fraction of checks passed, detector-scored): 0.55 overall — close to half of generations miss a requested attribute, the weakest of DALL·E 3 / FLUX / SDXL; counting and spatial relationships are the specific weak points. Independent (GenEval paper, Ghosh et al. 2023; echoed in later comparison tables). Structured score in benchmarks below.

Text rendering Weak

A well-known weakness of SDXL-era models.

Resolution ceiling Moderate

Native 1024×1024; upscale or tile for larger.

Inference speed Moderate

Multi-step; per-second compute cost depends on hardware and step count.

Benchmarks

Benchmark Score Config Source
GenEval 0.55 SDXL overall GenEval as reported in the GenEval paper; also echoed in later comparison tables. paper ↗

Operator guidance

Route here when you need open weights, fine-tuning/LoRA control, or self-hosting economics. Step up to DALL·E 3 or FLUX for prompt adherence and in-image text; SDXL is the weakest of the three on text and complex composition.

Use cases

Provenance notice

  • Open weights under a RAIL licence carrying use-based restrictions
  • No built-in content-provenance metadata by default

Limitations

Citations