SDXL¶

Development line: project:sdxl · thread sdxl-development
Last event: 2023-06-22 · 2 dated since 2023-05-24 · Researched: 2026-09-04 · confidence: medium

What it is¶

SDXL: a Stable Diffusion model for practitioners who need prompt-driven image generation.

• Generates text-to-image, image-to-image, and inpainting outputs.
• Runs the official 1.0 base checkpoint alone or with an optional refiner.
• Default target: 1024×1024; default checkpoints are not recommended below 512 pixels.

Development line¶

• 2023-05-24 — SDXL training-progress milestone was reported. On 2023-05-24, a dated SDXL development link indicated that model training had reached a midway milestone and pointed readers to a community discussion channel. This is a material checkpoint because it records the project in an active pre-release training and feedback stage.
• 2023-06-22 — Stability AI published an SDXL 0.9 update. It was available through Clipdrop, research weights were available, and 1.0 was planned for mid-July.

What changed¶

2023-05-24 — SDXL was still about half trained; a Discord bot collected image-preference votes while testing parameters, so this was training feedback rather than a release. 2023-06-22 — SDXL 0.9 moved from beta toward a limited research release: it was available through Clipdrop, research weights were available, and 1.0 was planned for mid-July. 2023-06-25 — the linked official post cannot be retrieved today, so no additional technical change can be verified from it. 2023-07-26 (found 2026-09-04) — SDXL 1.0 was released as an open model, replacing the limited research-only 0.9 release with the base-plus-optional-refiner 1.0 line. 2026-09-04 (found today) — the maintained practical path is the official stabilityai/stable-diffusion-xl-base-1.0 checkpoint through Diffusers; the base remains usable without the refiner.

How to use this¶

From 2023-06-22, practitioners should track SDXL 0.9 separately in model and workflow notes: this line moves from a training-status update to a named SDXL 0.9 publication.

1. Install Diffusers, Transformers, Accelerate, and the optional invisible-watermark dependency from the current SDXL guide. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
2. Load stabilityai/stable-diffusion-xl-base-1.0 with StableDiffusionXLPipeline or AutoPipelineForText2Image, preferably as fp16 safetensors on a supported accelerator. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
3. Start text-to-image work at 1024×1024; use 768×768 or 512×512 only when the workflow needs it, and avoid smaller dimensions with the default checkpoint. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
4. For an edit, reuse the loaded pipeline as image-to-image with an input image, or as inpainting with an input image and mask. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
5. Add stabilityai/stable-diffusion-xl-refiner-1.0 only when a final denoising pass is worth the extra compute; the documented ensemble handoff is 80% base denoising and 20% refiner denoising. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
6. Before publishing or automating output, check the model card's license, intended-use boundary, and known failures for text, faces, composition, and factual depictions. — https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

Best practices¶

• Use the official 1.0 base checkpoint as the baseline and treat the refiner as optional; the base model is a supported standalone path. — https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
• Keep default SDXL generation near its 1024×1024 native target; use size and crop conditioning for composition rather than expecting the default checkpoint to perform well below 512 pixels. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
• For out-of-memory errors, use model CPU offload before sacrificing the documented resolution range; use torch.compile or xFormers only when the runtime supports them. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
• Keep the invisible-watermark dependency when generated-image marking is required by the workflow; Diffusers enables it by default when installed. — https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl
• Do not use SDXL output as a factual representation of people or events, and retain human review for legible text, faces, and spatial relationships. — https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

Superseded by this¶

• 2023-05-24 — Discord voting and parameter experiments were a pre-release training-feedback process, not a deployment path; SDXL 1.0 open weights superseded that state on 2023-07-26.
• 2023-06-22 — SDXL 0.9 research-weight and Clipdrop-preview guidance is obsolete for a new deployment; use the named SDXL 1.0 base checkpoint instead.
• 2023-06-22 — The planned mid-July move to 1.0 is no longer future guidance; the open SDXL 1.0 release occurred on 2023-07-26.

Still unknown¶

• The supplied 2023-06-25 Twitter URL returned no readable post data in the available reader; its exact claim and technical impact remain unverified.
• This does not assert current availability, pricing, or support status for the legacy Stability API engines named in 2023; the current evidence supports the checkpoint-and-Diffusers workflow only.

Sources¶

Agent brief¶

• Subject: project:sdxl, thread sdxl-development, 2 dated events 2023-05-24 → 2023-06-22.
• Practical note: From 2023-06-22, practitioners should track SDXL 0.9 separately in model and workflow notes: this line moves from a training-status update to a named SDXL 0.9 publication.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.