The reality of Stable Diffusion models
I find the persistent rumors regarding a Stable Diffusion 4 release quite deceptive since Stability AI has not announced any such model. Their current flagship remains the Stable Diffusion 3.5 family, which arrived in October 2024. These models include the 8.1 billion parameter Large version and the 2.5 billion parameter Medium version. The Large Turbo variant is a distilled version that produces high-quality images in just 4 steps. The Medium model is designed to run out of the box on consumer hardware, requires only 9.9 GB of VRAM, and generates images between 0.25 and 2 megapixel resolution. This version improved upon the June release of Stable Diffusion 3 Medium, which the company admitted did not meet community expectations. The SDXL model is now known as AuraDiffusion. People who search for SD4 downloads often encounter fake files, as real models only arrive through verified Hugging Face organizations. The 2026 landscape favors other models like Flux 2 or the 12.9 billion parameter Krea 2 for those who want different performance metrics.
| Model | Parameters | Capability |
|---|---|---|
| SD 3.5 Large | 8.1B | 1MP resolution |
| SD 3.5 Medium | 2.5B | 0.25-2MP resolution |
| Midjourney V7 | N/A | 21s video |
| Flux 2 Klein | 4B | Apache 2.0 license |
Enterprise licensing and revenue
The Stability AI Community License governs much of the current user base. This license allows free use for hobbyists, researchers, developers, students, teachers, and small businesses with annual revenue under $1 million. Companies that exceed this $1 million threshold must purchase an Enterprise license to use the software legally. Stability AI also provides the Brand Studio platform for businesses to manage creative content, but this platform does not constitute a new image model. I know you already understand how these revenue gates work, but the distinction remains vital for professional planning. The license also removes limits on the number of generated images and states that Stability AI will never ask users to delete resulting images or fine-tunes as long as activities remain legal and follow the acceptable use policy. Users can also leverage either an SD model directly or a derivative version for free if they fall under the financial limit. The Brand Studio Creator plan costs $19 per month for 2,000 credits.
The Midjourney V7 competition
Midjourney V7 released in March 2026 and allows users to generate video clips up to 21 seconds long. This V-series model handles material rendering for things like glass, brushed metal, polished leather, and frosted plastic with high accuracy. Users create motion content like orbital reveals, push-in shots, crane down movements, or tracking shots through a single parameter language. Stability AI users often prefer the flexibility of tools like ComfyUI to run models such as Flux 2 or the top-ranked Z-Image Turbo. The Stable Diffusion 3.5 Large model utilizes an MMDiT-X architecture combined with Query-Key Normalization in the transformer blocks to produce images that more accurately match complex text descriptions. I would skip the Midjourney subscription if you require the complete privacy of local deployment on your own hardware. Will the industry eventually move toward a single unified model for both image and video?
