SDXL: What It Is, What It Needs and How It Compares

SDXL, short for Stable Diffusion XL, is a text-to-image model released by Stability AI in July 2023 that generates at a native 1024x1024 resolution, using a larger network and a second text encoder than earlier Stable Diffusion versions. It is free to download under the CreativeML Open RAIL++-M licence, Stability AI says it should work on consumer graphics cards with 8GB of VRAM, and newer families such as SD 3.5, FLUX and Z-Image now sit beside it.

Checked 13 October 2026 against the SDXL paper; the base and refiner model cards, their file listings and the licence file; Stability AI's SDXL 1.0 announcement, generative-models repository, SD 3.5 announcement and licence page; the model cards for SDXL Turbo, SD 1.5, SD 3.5 Large, FLUX.1 [dev], FLUX.1 [schnell], FLUX.2 [dev], FLUX.2 [klein] 4B and Z-Image-Turbo; and ComfyUI's SDXL examples, FLUX.1 tutorial and README. We publish no speed or quality benchmarks here because we have not run any.

What is SDXL?

SDXL followed the 1.x and 2.x versions of Stable Diffusion. Its paper, submitted on 4 July 2023, says it "leverages a three times larger UNet backbone" than previous versions and adds a second text encoder. Stability AI's repository records the release of SDXL 1.0 on 26 July 2023, and the announcement gives the sizes: a 3.5 billion parameter base model, and 6.6 billion parameters for the base and refiner together.

It comes as two models. The model card describes the base as generating the latents, which the refiner, "specialized for the final denoising steps", then finishes. The card adds that "the base model can be used as a standalone module", and that the two-stage route is "slightly slower". The second stage is the technique covered in our img2img guide.

SDXL hardware requirements

Stability AI's announcement says SDXL 1.0 "should work effectively on consumer GPUs with 8GB VRAM or readily available cloud instances". That is the only hardware figure the publisher gives. On disk, sd_xl_base_1.0.safetensors is 6.94 GB and sd_xl_refiner_1.0.safetensors is 6.08 GB. The table sets that beside what other publishers state for their models.

None of these is a speed test. A model that loads on a small card can still be slow, and the project behind ComfyUI publishes no generation time per card. The requirements ComfyUI itself documents are on our ComfyUI statistics page.

SDXL resolution, the refiner and SDXL Turbo

Resolution is the setting that most often goes wrong. ComfyUI's SDXL examples say that for optimal performance "the resolution should be set to 1024x1024 or other resolutions with the same amount of pixels but a different aspect ratio", and name 896x1152 and 1536x640 as good ones. The 512x512 habit from Stable Diffusion 1.5 does not carry over.

SDXL Turbo is a separate, distilled model. Its card says it samples in 1 to 4 steps, does not use a guidance scale or a negative prompt, and preferably generates at 512x512. It also has a different licence from SDXL 1.0, which matters in the next section.

The SDXL licence

SDXL 1.0 is released under the CreativeML Open RAIL++-M licence, dated 26 July 2023. Three parts of the text matter most to someone publishing images.

• The grant. A "perpetual, worldwide, non-exclusive, no-charge, royalty-free" licence to reproduce and distribute the model and its derivatives. The text contains no non-commercial clause.
• The output. "Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses."
• The restrictions. Attachment A lists uses you agree not to make. They include breaking any applicable law, exploiting or harming minors, spreading verifiably false information to harm others, and defaming, disparaging or harassing others.

Those restrictions travel with the model. The preamble says derivative versions "will always have to include", at minimum, "the same use-based restrictions as the ones in the original license", which covers a fine-tuned checkpoint or a LoRA built on SDXL.

SDXL vs SD 1.5 vs SD 3.5 vs FLUX

Here is what each publisher states about its own model today. If you want the two most-compared families side by side, see our FLUX vs SDXL comparison.

The pattern is clear enough without a benchmark. Among the models with a stated parameter count, only SD 3.5 Medium is smaller than the SDXL base, and SDXL is the only one under the Open RAIL++-M licence. The newer families are larger, and they split into two licence groups: Apache 2.0 for FLUX.1 [schnell], FLUX.2 [klein] 4B and Z-Image, and revenue-capped or non-commercial terms for SD 3.5 and the FLUX [dev] models.

The SDXL model card is frank about limits. It says the model "does not achieve perfect photorealism", "cannot render legible text" and struggles with compositional prompts, and that faces and people "may not be generated properly". Text in the image is one place the newer families advertise progress: the Z-Image team, for example, lists accurate English and Chinese text rendering among its strengths.

Is SDXL still worth using in 2026?

For many persona workflows, yes, and the reason is the files built around it. Add-ons belong to one model family, and SDXL has had years of them.

Three add-ons do most of the work for a persona on SDXL. A character LoRA holds the face, and the free LoRA training steps calculator plans the run. ControlNet holds the pose. InstantID carries a face from one reference image, with licence limits of its own that you should read first.

Whatever model you pick, the same rule applies to faces. The SDXL card says the model "was not trained to be factual or true representations of people", and its licence forbids using it to defame or harass. A persona is a fictional adult, and reproducing a real person's face without their written permission is not allowed on any model. The AI Influencers program teaches the full workflow on current models, from base-model setup to a trained face and the disclosure rules for publishing.

SDXL FAQ