What Is Stable Diffusion — Open-Source Image AI: How It Works, Running Locally, and Commercial Licensing

On August 22, 2022, the London startup Stability AI released the weight file for an image generation model called Stable Diffusion v1.4. A single 4GB `.ckpt` file. The moment it hit GitHub and Hugging Face, "image generation AI" went from something behind the cloud to software you could download to your own PC. Neither Midjourney nor DALL·E 2 would do that at the time.

Almost four years on, Stable Diffusion has reached SD 3.5 Large (8.1 billion parameters), and Civitai hosts over 100,000 custom models and LoRAs. Meanwhile, the licensing blowback around SD3's release caused a developer exodus, which gave birth to FLUX — built by the original SD team's new company, Black Forest Labs — and FLUX has overtaken the parent in quality. The picture is no longer simple.

My stance up front. If "Midjourney is fine" works for you, don't force yourself into Stable Diffusion. But if any of these apply — "I want to keep the same character consistent across 100 images," "I want to mix in my own confidential data locally," "I want my monthly cost to be $0," "I need an open model I can disclose for commercial work" — then SD is unavoidable. This article covers how SD works, its version history, hardware requirements, licensing, ecosystem, and how to choose, all as of May 2026.

1. August 22, 2022 — The Day Image AI Became Something You Could Download

At the time, the image generation AI scene was a two-horse race: OpenAI's DALL·E 2 (invite-only beta) and Midjourney V3 (Discord-only). Both were cloud-only, and both kept their weights completely hidden. What their AI learned, how it ran, what it could and couldn't generate — all of it was at the vendor's discretion.

Then Stability AI made a choice nobody expected: release the weight file itself. A diffusion model trained on LAION-5B (5.8 billion image-text pairs), inference code under MIT, weights under CreativeML Open RAIL-M (commercial use OK, almost completely free). Within a week, engineers worldwide had it running in Google Colab, a local WebUI (later AUTOMATIC1111) was born, Civitai launched — and the personalization of AI art took off.

The remarkable thing wasn't the technical leap so much as the precedent: "image generation AI is something individuals can own and modify." If you want an LLM analogy, the shock was close to Llama 2 and Llama 3 dropping with "commercial use OK." Ever since, the image AI industry has run two parallel tracks: "closed and high quality" (MJ/DALL·E) and "open and freely customizable" (the SD family).

2. What Is Stable Diffusion — In Three Lines

Stable Diffusion is an open-weight, diffusion-model-based image generation AI released by Stability AI. Three-line breakdown:

The thing I think actually matters is what "diffusion model" means in plain terms. In the GAN era (StyleGAN and friends), a generator and a discriminator fought each other to produce images. Diffusion models took a different path: "start from a noisy image and gradually subtract noise." A simpler idea — but it turned out to produce far more stable, high-resolution output than GANs. That insight is the core of SD's success, and almost every image AI since (Imagen, DALL·E 3, FLUX) is also a diffusion model.

3. Version Lineage — SD1.5 / SDXL / SD3.5 and the FLUX Split

The most confusing thing about SD's history is "which version should I actually use?" Each generation differs in performance, license, recommended GPU, and LoRA ecosystem. Let's lay it out.

Bottom line: if you're starting today, it's a two-way pick between SDXL and FLUX.1 dev. SD 1.5 is light and has the most LoRAs, but it's a generation behind on quality. SD 3.5 Large is heavy yet pushed around by FLUX. The practical sorting is: SDXL for commercial design, FLUX for top quality, SD 3.5 Medium for the lightest viable local setup.

FLUX's arrival has an ironic backstory. After the SD3 licensing fiasco (more below), much of the original SD team left Stability AI, set up Black Forest Labs in Germany, and launched FLUX.1. "A higher-quality SD successor" — coming from the people who built SD in the first place. From the community's perspective, plenty of people now see FLUX as the rightful heir rather than the parent.

4. The Reality of Running It Locally — By VRAM Tier

"Runs locally" is one thing; what your specific PC can actually do is another. Here's what I've seen in practice.

No sugarcoating: if you want to seriously touch SD today, the realistic entry points are an RTX 3060 12GB (around $200 used) or an RTX 4070 (around $600 new). 8GB GPUs work, but you're walking into a swamp of optimization flags and quantization — not what I'd recommend to a beginner. If you don't want to buy a GPU, the right move is cloud inference services (Runpod / Replicate / Civitai's own hosting) at roughly $0.001–$0.01 per image.

5. The License Trap — Lessons from the SD3 Backlash

"It's open source, so commercial use is fine" is not the simple statement people want it to be with SD. The license depends on the version.

When SD 3 dropped in June 2024, its license was so harsh — usage-based fees per generated image, a ban on Civitai distribution of derivatives — that Civitai publicly refused to host SD3 derivatives. The community declared "SD is dead," many developers walked to Black Forest Labs and shipped FLUX. Stability AI massively loosened the terms when SD 3.5 launched in October (the current $1M revenue version), but as of May 2026, community trust has not fully recovered.

Practical advice: "Just use SDXL" is the version that bites least. CreativeML Open RAIL-M means no revenue cap, the LoRA pool is huge, and the ecosystem is mature. Move to SD 3.5 or FLUX only when SDXL stops being enough.

6. Civitai / LoRA / ComfyUI — An Ecosystem Bigger Than the Model

Talking about Stable Diffusion as "just the model" misses the point. SD's strength is the surrounding ecosystem.

One caveat. SD 1.5 LoRAs don't load on SDXL; SDXL LoRAs don't load on FLUX. Each base model is its own ecosystem. If the LoRAs you love on Civitai are all SD 1.5, switching to SDXL means abandoning them. When searching on Civitai, always check the "Base Model" filter.

7. Midjourney vs Stable Diffusion — Which to Pick

People often ask "which is better, SD or Midjourney/DALL·E?" — but that's the wrong axis. Go with Midjourney for quality, go with SD for freedom and ownership. Different roles entirely.

The clean read: for "make a single pretty image," Midjourney. $10/month and no setup hell. For "I want 100 images of the same character," "I want to mix in proprietary data," "I want a commercial flat-rate at any volume," or "I want to reproduce a specific anime style," Stable Diffusion. Neither is "better." Plenty of pros use both (an illustrator I know roughs out composition in MJ and finishes in SD).

8. Three Pitfalls — Copyright, NSFW, Compatibility

Three things you'll hit using SD that are worth knowing up front.

Pitfall ①: Training-data copyright risk

SD's base models are trained on LAION-5B (5.8 billion images scraped from the internet). Inevitably, copyrighted works are in there in large numbers. Getty Images is currently suing Stability AI (filed 2023, ongoing in both US and UK), and "specific artist style" LoRAs on Civitai have gotten visibly greyer since 2025. For commercial work, minimum hygiene: don't prompt by specific artist names, and even on Civitai LoRAs, avoid public figures or works modeled on identifiable copyright holders. If "commercial safety" is non-negotiable, Adobe Firefly is the alternative.

Pitfall ②: NSFW generation is trivially easy

Because SD has open weights, disabling the SafetyChecker means sexual or violent images are easy to generate. Civitai openly hosts many NSFW models. The technology itself is neutral, but creation or distribution of generated content involving minors is illegal in many countries (Japan currently has legislation under discussion). Never do this on a work PC during work hours — logs and network traffic make it trivial to spot. Even on a home PC, certain categories are illegal to create or even store. Self-awareness is mandatory.

Pitfall ③: Generational compatibility splits

As covered above, SD1.5 / SDXL / SD3.5 / FLUX are each their own ecosystem. LoRAs, embeddings, and ControlNet models don't cross-load. "Let me upgrade to SDXL" can mean discovering 50 SD1.5 LoRAs you can't use anymore. If you're starting out, pick one (SDXL or FLUX) and stay within that ecosystem — it's actually more efficient in the long run.

Summary

Stable Diffusion changed the world in 2022. But in 2026, "just use SD" is no longer the default answer — Midjourney V8 wins on raw quality, Adobe Firefly wins on commercial safety. The reason SD hasn't died — and in fact has gained momentum with FLUX — is that it remains the only option for "use image AI on your own PC, with your own data, exactly the way you want, without depending on any cloud company." Midjourney can lock you out of Discord; OpenAI can change its terms of service; the SD weight file on your SSD is yours. For people who feel safer that way, SD will keep being a special tool.

FAQ