By Stability AILearning Objectives
- Understand what Stable Diffusion 3.5 is and how its open-weight approach differs from proprietary image generators
- Compare the three model variants and their hardware requirements
- Assess Stability AI's current financial health and the Getty Images lawsuit outcome
What Is Stable Diffusion 3.5?
Stable Diffusion 3.5 (released October 2024) is the latest generation of Stability AI's open-weight image generation model. Unlike DALL-E or Midjourney — which are only accessible through APIs or web interfaces — SD 3.5 model weights are publicly available, allowing anyone to download, run locally, and fine-tune the model for their specific needs.
The model uses an MMDiT (Multi-Modal Diffusion Transformer) architecture with improved text rendering — one of the weakest areas of previous Stable Diffusion versions.
Model Variants
| Variant | Parameters | VRAM Required | Best For |
|---|---|---|---|
| Large | 8.1 billion | ~18 GB (11 GB with FP8 quantization) | Highest quality; best prompt adherence |
| Large Turbo | 8.1 billion | ~24 GB (reducible with quantization) | Speed-optimized 4-step generation |
| Medium | 2.5 billion | ~10 GB | Consumer GPUs (RTX 3080+); balance of quality and speed |
License
| Use Case | License | Cost |
|---|---|---|
| Individuals | Stability AI Community License | Free |
| Businesses (under $1 million annual revenue) | Stability AI Community License | Free |
| Businesses (over $1 million revenue) | Enterprise License | Paid (contact Stability AI) |
Model weights are available on Hugging Face. This is not a traditional open-source license (no OSI approval) but is permissive for most users.
Getty Images Lawsuit
In November 2025, the UK High Court rejected Getty Images' central copyright claim against Stability AI — finding only limited trademark infringement. This was the first major UK ruling on copyright and generative AI training, though Getty's CEO signaled continued legal pressure.
SD 3.5 vs. Competitors
| Model | Strength | Key Difference |
|---|---|---|
| Stable Diffusion 3.5 | Free; self-hostable; fine-tunable; runs on consumer GPUs | Open weights; maximum customizability; community ecosystem |
| DALL-E 3 | Best out-of-box quality for non-technical users | API-only; no local deployment; higher per-image cost |
| Midjourney | Strongest aesthetic quality; large community | Proprietary; no self-hosting; no fine-tuning |
| Flux (Black Forest Labs) | Strong open-weight competitor; excellent prompt adherence | Some benchmarks show Flux Pro exceeding SD 3.5 |
| Ideogram | Best text-in-image rendering | Niche strength; less customizable than SD 3.5 |
Company Details
| Detail | Info |
|---|---|
| Company | Stability AI |
| CEO | Prem Akkaraju (appointed June 2024; former CEO of Weta Digital) |
| Headquarters | London, UK |
| Estimated Revenue (2024) | ~$50 million |
| Website | stability.ai |
Key Takeaways
- SD 3.5 is the leading open-weight image generation model — free for individuals and small businesses, with three variants for different hardware
- Medium variant (2.5 billion parameters) runs on consumer GPUs (RTX 3080+) with ~10 GB VRAM — no cloud required
- UK High Court rejected Getty's central copyright claim (November 2025) — a landmark ruling for generative AI
- Best for users who want maximum control: local deployment, fine-tuning, and customization without per-image API costs
