Stable Diffusion¶

Development line: project:stable-diffusion · thread stable-diffusion-development
Last event: 2022-07-26 · 1 dated since 2022-07-26 · Researched: 2026-09-04 · confidence: medium

What it is¶

Stable Diffusion is a family of generative image models for artists and developers who need a choice between self-hosting and an API instead of a closed service like Midjourney.

• text-to-image and image-to-image
• ControlNet for control via Canny, depth or blur
• Large, Turbo, Medium and Flash variants

The SD 3.5 API outputs 1 MP images, default 1024×1024; image-to-image input accepts up to 10 MiB.

For new integrations, use SD 3.5 rather than earlier SD3 or SDXL beta identifiers.

Development line¶

• 2022-07-26 — Stable Diffusion beta-access signup milestone. On 2022-07-26, the Stable Diffusion record linked to Stability AI's beta-signup form. This supports retaining a beta-access milestone, but it does not establish a released model version, eligibility, or feature set.

What changed¶

2022-07-26 — beta registration meant limited testing rather than public weight release; research access opened 2022-08-10, public release followed on 2022-08-22.

2023-03-09 — the NeuralPic link specifies no dataset, model or Stable Diffusion release; product change unconfirmed.

2023-03-25 — the stable_diffusion_v2_2_2_xl_beta entry has no URL, so the date and change cannot be verified. On 2023-04-13 the official API added the stable-diffusion-xl-beta-v2-2-2 engine for SDXL beta and DreamStudio.

2024-03-07 — the link points only to an X profile, leaving the specific post unverifiable. The official status of SD3 in this window was early preview and waitlist with variants from 800M to 8B parameters.

2024-03-13 — the specific X post is unavailable to check; no confirmed model or API change comes from it.

On 2024-06-12 SD3 Medium was released for REST API and weight download; on 2024-10-22 SD 3.5 Large and Large Turbo appeared, and on 2024-10-29 SD 3.5 Medium followed.

From 2025-04-17 SD3 API identifiers redirect automatically to SD 3.5 equivalents. In the active Core Models list from 2026-05-20, SD 3.5 Medium, Large and Large Turbo remain.

How to use this¶

For the 2022-07-26 historical record, treat beta signup—not a confirmed public release—as the operational access route; confirm current availability and terms separately.

1. Choose a path: API for managed integration or local hosting for infrastructure control. For 1 MP quality take Large, for speed pick Large Turbo or Flash, and for lower resource cost pick Medium.
   — https://platform.stability.ai/docs/api-reference
2. For the API, create an account and key, then pass the key only in the Authorization header.
   — https://platform.stability.ai/docs
3. Send a multipart POST to /v2beta/stable-image/generate/sd3 with prompt, model and output_format; for image-to-image pass the source image, strength and mode=image-to-image.
   — https://platform.stability.ai/docs/api-reference
4. For local runs, accept the weight license, install diffusers, transformers and accelerate, load StableDiffusion3Pipeline, and save output images.
   — https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
5. To anchor composition to reference images, attach SD3.5 Large ControlNet with a Canny, depth or blur condition image.
   — https://github.com/Stability-AI/sd3.5

Best practices¶

• Do not store or commit API keys: pass the key in the Authorization header and rotate it if leaked.
  — https://platform.stability.ai/docs
• Set an explicit seed for repeatable comparisons: it controls generation randomness.
  — https://platform.stability.ai/docs/api-reference
• For SD 3.5 Medium, keep prompts within 256 T5 tokens: longer requests risk edge artifacts. Use Skip Layer Guidance for structure and anatomy.
  — https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
• When VRAM is tight, use official 4-bit quantization and CPU offload instead of arbitrary weight swaps.
  — https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
• Check the license before commercial self-hosting: the Community License covers entities under $1M annual revenue, above which an Enterprise License is required.
  — https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
• Add custom content-safety guardrails in production: training filters do not prevent harmful outputs.
  — https://huggingface.co/stabilityai/stable-diffusion-3.5-medium

Superseded by this¶

• 2022-07-26: joining the beta waitlist was superseded by the 2022-08-22 Stable Diffusion public release.
• 2023-04-13: engine limits and the name stable-diffusion-xl-beta-v2-2-2 apply to historical SDXL beta, not current API routes.
• 2024-02—03: waiting for SD3 early preview was superseded by SD3 and SD 3.5 releases.
• 2025-04-17: API instructions targeting sd3-large, sd3-large-turbo and sd3-medium are obsolete because requests redirect to SD 3.5.

Still unknown¶

• Content of the 2022-07-26 beta registration form is unavailable; its role as preliminary access rests on later official announcements.
• The NeuralPic link from 2023-03-09 does not connect the event to a specific Stable Diffusion dataset, checkpoint or release.
• The 2023-03-25 record lacks a URL: the name resembles the later SDXL beta API engine, but does not prove it appeared on that date.
• X links from 2024-03-07 and 2024-03-13 returned no content, leaving model, license, or management changes unverified.
• Tags dataset, stable_diffusion and stable_diffusion_v2_2_2_xl_beta may conflate an unconfirmed dataset link, an unofficial beta lead, and official model releases.

Sources¶

Agent brief¶

• Subject: project:stable-diffusion, thread stable-diffusion-development, 1 dated events 2022-07-26 → 2022-07-26.
• Practical note: For the 2022-07-26 historical record, treat beta signup—not a confirmed public release—as the operational access route; confirm current availability and terms separately.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.