API for generating images from text. Rate my idea
The open-sourced Stable Diffusion model is becoming the next revolutionary platform for developers and creators. Generating images from text input in a couple of seconds enables infinite possibilities for building next-gen apps, breakthrough products, and creative experiences.
But using and running the model in a scalable way is a hard, expensive, time and resources consuming. Most of developers and creative people, do not want to spend time doing that. They want to focus on creativity, ideas and building products, not on scaling the GPU infrastructure.
The idea is to give developers simple-to-use API for generating images based on text prompts. My next SaaS product (getimg.ai) will be about that. I created a simple website, where you can see working examples, check out prices and sign up for the waiting list.
What do you think about the idea, would you use and pay for such a product?
Check the website here (https://getimg.ai) 👈🏻
Any feedback is welcome!
Just a simple stable diffusion API is not enough. As others said, there is a LOT of competition, as stable diffusion is open source.
If you want to keep this project as API-only, I would think about adding a lot more features to differentiate from competitors (support for large resolutions, more models, img2img, upscaling, inpainting/outpainting, bg removal, face correction, loopback, prompt weighting, finetuning) and offering blazing-fast generation speed (< 5 sec for each request, top-end GPUs can achieve that). Basically, you'd have to niche down and build the go-to API for all this AI art stuff.
Another option is to build an AI art generation platform/social network, like dalle2, nightcafe, dreamstudio, pixelz, etc, and offer API as a secondary thing. I think this path is more sustainable and has more potential.
I am actually building an AI art generation platform of my own, dreamlike.art. I would be happy to use an already-made API, but there are none that are cheap and offer all the features I want. That's why I'm building it all from scratch.
Thanks for your thoughts! I've seen dreamlike.art earlier today and I really like it.
I have a lot of the same ideas for features or pivots, but instead of building all of them I wanted to verify pricing first and see if people would pay for the simplest feature which is text to image. Not gonna lie, it's been hard, most people sees Stable Diffusion as a toy, not a real business opportunity.
The pricing is expensive because of costs. To deliver really fast responses (~2.5 sec) you need A100 GPU which is 20x more expensive than RTX 3090, then you need to make sure that hundreds of requests can be run parallel, and finally cover the cost of storage and bandwidth.
I see Stable Diffusion as a tool for artists, just like Photoshop. That's who I'm trying to target. With the help of AI artists can greatly speed up their workflow, they just don't know how yet, as it's all pretty new. As people get more familiar with AI assisted art production, demand will rise.
You can do 2 tiers: one with A100s, another with cheaper GPUs. People who need the extra-fast generation will pay for it. Anyway, getimg is 10 times more expensive than dream studio. I don't think it's competitive at this price. 5 sec is good enough for most people. I personally wouldn't pay 20x for 2x speed up.
Thanks for the suggestions!
It's 10x more expensive than a dreamstudio, but only for the requests with default params. I did not want a complicated pricing, that charges more for higher resolution, more steps etc. Anyway I need to revise the idea and business model a bit.
I think it's fine if it's complicated, it's an API after all. It doesn't have to be super simple. I would actually want more control.
Also apparently people really, REALLLY like generating porn. Like a third of the requests are porn related. And not normal porn, but some weird stuff.
Do you use NSFW filters on dreamlike.art?
I find it triggering on some images that I would not consider NSFW. I myself use the website, so it's disabled for now. I'm thinking about enabling it for the free plan when I add the paid plans, and adding a toggle in the settings for the paid plans.
Can I ask what is your differentiator from Hugging Face?
Sure. Hugging Face does not provide Stable Diffusion API (it does offer Inference API for other AI models, but large models are supported only by Enterprise plans with yearly commitment). It is where Stable Diffusion model is available to download, not an interface to interact with.
getimg.ai offers scalable API and will also host and serve generated images.
I won't lie, I am having trouble seeing the business. You are making compute+stable diffusion available through an API with a side of hosting, right?
Compute is cheap, stable diffusion is free. The only main thing is the convenience of an API. It might work but I'd assume fierce competition. That's what happens when you use other's IP. Barriers to entry is very low. Might easily become a race to bottom with everyone cutting prices. I hope you have a plan around that.
Similar thoughts here. In a new tech frenzy like this, everyone runs after building platforms. no one is building apps. You should focus on building the killer app on stable diffusion. I would niche down to some very tiny use case where stable diffusion could actually solve a problem. Don't try to chase a generic art generation tool - that's obvious and everybody will try to do that.
Instead why not try something like - (note this is just a hypothetical example of niching down) - Build an app that lets travel photographers add subjects into landscape shots to create a certain mood. my point is, get really specific because your value-add is going to be the application layer on top of stable diffusion, not just giving people access to the open source api.
I agree that more people should focus on building killer apps, not on generic websites/platforms.
All niche down ideas I've got require generating multiple images in parallel fast, hence the idea - build scalable API first (it's not as easy as it sounds), give access to the API to others, focus on building apps with the API.
That's fair. I guess the challenge is with keeping this experimentation phase sustainable for you. If people are willing to pay for the API, you can fund the development of whatever you want to build. Another way to think about this then is in the quality of the API itself. If you can offer some flavor of the API that feels different of just document the hell out of it so well that using your API becomes way easier than using anyone elses.
I do think that a simple API over an open-source library will lead to a lot of competition, and likely a free or super cheap product would win out.
I put in "Business ideas for generating images with AI" into bizideas.ai and got a few decent results. Maybe something a bit more niche could help you. Good luck!
Will it be AI generated images or a simple google search with higher accuracy?
All images will be AI generated. All examples on https://getimg.ai are AI generated too.
Interesting concept for sure. I'm glad I read because at first i just thought this was another dall-e ! I think this could be useful, especially for something like blog posts where you don't need any specific sort of image but want something relevant and royalty free
There are already, so many interfaces similar to Dall-E, for playing with Stable Diffusion model. There's no sense in building another one, especially, that there are mostly free for now.
For sure, but i agree that what you're doing is quite different!
great start
Thanks!
I think you have a great idea in a growing market, have you defined yet what makes your ideal customer profile?
Combine this with a model that has been trained to churn out high quality images in line with the branding guidelines, and you have an amazing tool for A/B testing and real time personalized browsing/shopping experience.
Congratulations for this great idea sir! :)
I just wonder what makes it different from Midjourney for instance.
It’s not supposed to be a software/interface for artists to generate images. For now just a simple API to embed images into websites/apps/web projects based on text queries encoded in the url.
Number of comments alone means you just validated your idea :)
Btw awesome idea, I wanted to build same with Dall-e but saw they don't provide API.
Thanks! I try to be realistic, it's not so obvious that it's been validated. I will probably write about my thoughts and the outcomes next week.
I have access to GPT-3 API to create those images. But I imagine if I didn't I'd use your service.
I think you meant Dall-E, GPT-3 is a text model. OpenAI's Dall-E does not offer API to generate images at scale, only a web interface
I’d be too lazy to set up a model on my own, so I’d use a service like this :)
If I were opening up the service to website visitors to use (eg as part of a trial or to allow them to interact with an app without signing up), pricing would be important, as well as features to limit generation by IP.
For reference, this is DreamStudio’s (stable diffusion api) pricing https://beta.dreamstudio.ai/membership. It’s currently charging £10 for 1000 generations.
Great idea! Limitation by IP, could be added.
As for comparison to DreamStudio's pricing, it's charging $10 per 1000 generation with default params (512x512px, 50 steps, etc.), but for example, if you want larger resolution, you'll pay up to 30 credits per one image.
getimg.ai does not intend to limit pricing based on parameters. Every generated image will be priced the same.
I was playing around with seeing if it was feasible to run stable diffusion on a server just for my own fun (I don't own a computer that is powerful enough to run it), I'd be interested in what you found. The cheapest VPS I could find that could run it was about $40 a month and it took about 8 minutes per image (default 32 steps). It definitely didn't have enough memory or cpu to handle multiple image renders at once.
The idea seems very straight forward, but I don't know how this idea scales to be handling many API requests at once without making the request wait an unreasonable amount of time. I would think the hardware needed to do this would be very expensive (at least for my budget).
I'd love to be wrong though, this is just the first thoughts that came to my head. Let me know if you find something that works.
Running Stable Diffusion on CPU instances does not make sense, you'll need a VPS with GPU. Yes, It is expensive. The fastest you can generate image is with NVIDIA A100 in about 2-3 seconds ($3.5 per hour). With other, cheaper cloud GPUs you can achieve 7-10 seconds per image.
With optimization it is possible to handle hundreds API requests at once.
Did you consider using vast.ai? It's much cheaper than ordinary VPS rental. A100 go for 1$/h here
No, never heard of it, was using Google Cloud. I will definitely give it a try, as it's 4x cheaper.
Thanks again for the suggestion @sviasem! I updated the pricing based on vast.ai costs
GOD! That would help me a lot withmy project small hand grips . Is there an app for this? for convenience indeed.
This comment has been voted down. Click to show.