Overview
- Tool name: Stable Diffusion 4 (SD4 Base / SD4 Ultra)
- Developer: Stability AI
- Official website: https://stability.ai
- Category: AI image models
- License: Stability AI Community License (open weights)
Stable Diffusion 4 launched on April 6, 2026 - the first full-generational update since SD3.5 and the first major output from Stability's restructured team after its 2024 leadership crisis. The headline architectural change: a full move off the U-Net backbone that powered SD 1.x through SDXL onto a diffusion transformer (DiT), with a mixture of local and global attention layers and RoPE positional encoding for spatial awareness.
Two tiers:
- SD4 Base: 12GB+ VRAM, up to 1024x1024, aimed at community fine-tuning (LoRA / Dreambooth) and cost-efficient self-hosting
- SD4 Ultra: 24GB+ VRAM, native 4096x4096 output - the only open-weight model reaching that resolution without tiled upscaling workarounds - plus a dedicated text glyph conditioning module, a first for the SD family, fixing diffusion's long-standing inability to render legible text
Internal benchmarks: anatomically correct hands in about 87% of samples at standard guidance (roughly 60% for SDXL, 72% for SD3.5); training incorporated a curated PBR dataset subset, improving specular highlights, subsurface scattering in skin and consistent shadow directionality.
License notes (frequently misread)
The Stability AI Community License has three tiers:
- Research: free for any organization, no revenue threshold
- Commercial: free for organizations under $1M USD annual revenue
- Above $1M: an enterprise license negotiated with Stability's sales team
That is a step function, not linear per-call pricing: once licensed, self-hosted marginal cost per image approaches zero. Stability AI Solutions - the managed enterprise tier - now offers SOC 2 Type II and SOC 3 compliance, indemnification and dedicated support, which is exactly what previously kept regulated industries (financial services, healthcare, insurance) away from open-weight image generation in production.
Who it is for
- Print and large-format output: SD4 Ultra's native 4096-squared removes an entire post-processing step
- Assets containing text: packaging, UI mockups, storefront signage, product labels - the text module makes in-image copy usable
- In-house brand engines: open weights plus LoRA/Dreambooth, trained on your own style with data staying inside your network
- 4K+ production pipelines: pairs with Stable Video Diffusion 4D 2.0 and Stable Virtual Camera as a full visual-generation platform
Limitations
- The free-commercial threshold is only $1M annual revenue, so most mid-size companies need a paid license
- Ultra needs 24GB+ VRAM - a real hardware bar
- Long or stylized typography still trails closed models like GPT Image
Frequently asked questions
What is the latest Stable Diffusion version?
SD4, released April 6, 2026, in Base (1024 squared) and Ultra (4096 squared) tiers, both open weights under the community license.
Is SD4 free for commercial use?
Research use is free at any scale; commercial use is free below $1M annual revenue, with an enterprise license above that (per Stability's official license terms).
Ultra or Base?
Ultra for native 4096-squared, text rendering and top quality (24GB+ VRAM); Base for throughput, low-cost self-hosting and community fine-tuning (12GB+ VRAM).
SD4 vs FLUX 3 vs Midjourney?
SD4 Ultra for open weights plus ultra-high resolution and text rendering; FLUX 3 Klein for 8GB single-card private deployment and precise layout control; Midjourney for out-of-the-box aesthetics and a managed service.
*This page is not affiliated with Stability AI. Specs and licensing verified as of October 2026; see stability.ai for the authoritative terms.*
