AI
Stability AI’s Model Lineup in Azure AI Foundry
Stability AI’s three models in Azure AI Foundry serve different production needs. Stable Image Core runs on top of SDXL and handles high-volume batch jobs, tens of thousands of images per hour. SD 3.5 Large is the current flagship, built for photorealistic output, high prompt adherence, and text accuracy. Stable Image Ultra layers on top of SD 3.5 and adds editing workflows, so teams can modify generated images in tools like Paint. All three deploy from the Azure model catalog under the Stability AI collection, with API documentation under the consumption section.
Prompt-Based Image Generation in Practice
The Azure AI Foundry image playground walkthrough (04:46) shows SD 3.5 Large generating crisp, detailed images in roughly 9 seconds from a 1.5-line plain-language prompt. No special syntax, no negative prompts, no weighted keywords.
“And the problem, if you noticed, is very natural language.” — Addie
The model supports 9 aspect ratios and probabilistic generation by default, so each run produces a different image. Setting a seed value locks the output for reproducibility. The demo moves through a mountain fashion shot, a diversity test with a 1970s-style portrait, and an aspect-ratio comparison, all from the same conversational prompt style.
Product Photography and Text-in-Image at Scale
Product photography that once required a full shoot can be replaced with a prompt. The demo generates Diwali-decorated headphones on a wooden table for the Indian market, no studio required. Wrapping target words in double quotation marks inside the prompt tells the model to render legible text. A prompt for a highway billboard reading “stability for everyone” produces an image with that phrase visible in the frame. The same technique applies to product launches: generate the headphone image, then add “Happy Diwali” as a wide shot to give the text enough canvas to display clearly.
Safety Filtering and Brand-Specific Fine-Tuning
Content filtering is baked into the model itself, not bolted on. Stability AI maintains a list of prohibited objects, banned famous-person likenesses, and NSFW categories that the model won’t generate.
“And we actually have a dedicated safety team.” — Addie
Azure adds a second layer: its own responsible AI content checks fire on top of Stability’s model-level filters. On fine-tuning, the demo session confirmed that brand-specific training is available now. A fine-tuned version of the model can generate actual Nike footwear rather than generic shoes. SD 3.5 also runs on ComfyUI for teams that need a local workflow outside Azure AI Foundry.
Q&A
How do Stability AI’s models handle safety in image generation at enterprise scale? A dedicated safety team maintains a model-level blocklist covering prohibited objects, famous-person likenesses, and NSFW content, with Azure’s responsible AI content checks running as a second layer on top. ▶ 10:48
Is this integrated into AI Foundry Local? No confirmed answer at time of recording, but SD 3.5 runs on ComfyUI for local workflows outside Azure AI Foundry. ▶ 12:36
Notable Quotes
Stable Diffusion is solving real world business efficiencies. Addie · ▶ 4:31
And the problem, if you noticed, is very natural language. Addie · ▶ 5:44
And we actually have a dedicated safety team. Addie · ▶ 11:11
Key Takeaways
- SD 3.5 Large delivers photorealistic images from plain-language prompts with no prompt engineering required.
- Azure AI Foundry hosts three Stability AI tiers, from high-speed batch generation to full workflow editing.
- Built-in safety filters plus Azure’s content checks block NSFW, famous-person likenesses, and prohibited objects by default.