Open image models move value from the model to the product

Stable Diffusion runs on a gaming card and is going open. Image generation becomes a cheap input, and the margin shifts to products that own the workflow.
A consumer graphics card with under 10 GB of memory can now turn text into a 512 by 512 pixel image in a few seconds. Stability AI released Stable Diffusion, the model that does it, to researchers on 10 August, with the weights hosted on Hugging Face and a public release promised soon.
Once a model of this quality fits on a gaming card, the marginal cost of an image falls toward the cost of electricity and the model itself stops being the moat. Value moves to the product built on top.
Two business models now compete for the same customer. Three weeks earlier OpenAI opened a DALL-E 2 beta to 1 million people, where every image is made on OpenAI's servers and $15 buys 115 extra credits, each one prompt returning four images. Closed providers sell generation by the credit and keep control of quality, safety and pricing. Open weights, which reached language models at GPT-3 scale in July, let any developer run, tune and embed the model without paying a toll per image, so applications built on them keep software gross margins instead of reselling someone else's compute.
Training cost has come into view too. Stable Diffusion was trained over one month on a cluster of 4,000 Nvidia A100 chips, using a subset of the open LAION-5B image dataset; the work was led by Patrick Esser of Runway and Robin Rombach of the CompVis group now at LMU Munich, with support from EleutherAI and LAION. A month on a few thousand A100 chips is a budget a well-funded startup can raise, not only a large corporation. Demand is already visible: more than 10,000 beta testers create 1.7 million images a day.
The opening is in products that treat the model as raw material and own a workflow: design and marketing tools, game concept art and textures, product photography, and models tuned on a customer's own images and style. Serving costs need a plan for both the cloud and the user's own device. Training data is a diligence item from day one, because the dataset behind these models was collected from the open web and image rights are not settled.
Openness is still a promise rather than a licence. Whether the public release ships the full weights, and what its licence allows for commercial use, decides how much of this margin builders keep. Closed image services now face a free model of comparable quality, and their price response will show how fast selling by the credit erodes. Courts carry the other risk: if artists or image libraries challenge training on scraped images, the cheap input acquires a legal cost. Whether Stability AI follows with the other open models it says it is training will show if this is one release or a steady supply.
Open weights turn image generation from a scarce service into a commodity input. Fanaura backs founders who build on the customer, the workflow and the proprietary data, where its thesis on machine learning companies points, not on access to a single model. When the model is free, the product is the business.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building creative or game tools on top of open image models? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.