Ask HN: Are there AI prompt builder products? | Hacker News Reader

Ask HN: Are there AI prompt builder products?

14 pointshanniabu13 comments

What I'm looking for is a tool that will help you create a good prompt. I'd specifically like this for images so I'll use that as an example, but it can be used for text, video, and procedures.

For example, I will write a prompt in my own words, then it'll walk me through a few questions to make it better, such as image color (colored, vibrant, neon, black/white, etc), what style image do I want (photorealistic, anime, drawing, digital, 3d, etc), lighting (soft, natural, direct, dim, etc), and whatever other details that might be helpful such as photography style, clothing style, scenery, etc depending on previous answers. I envision each of these steps having an example image for each option to give you an idea of what that means (like what does soft vs natural vs direct lighting look like.)

If this doesn't exist, I do believe this is a very useful product worth building.

View on HN

Logitech just build it into mouse.

https://www.logitech.com/en-us/software/logi-ai-prompt-build...

https://blog.logitech.com/2024/04/17/looking-to-improve-your...

huh this wasn't posted on April 1

That's a great example what I'm referring to but for content generation

It doesn't guide you but DALL-E 3 via ChatGPT essentially mimics an InstructPix2Pix workflow so you can continue to iterate on the same image requesting changes to the picture using simple English, i.e.:

Generate a picture of a house

house.jpg

Let's change it to colonial style

House2.jpg

I don't like the color let's change it to red.

House3.jpg

Redo it using acrylic oil painting.

Etc.

Since we're focusing on images right now, you can't just "build one prompt" to rule them all. Effective prompts are highly dependent on the checkpoint, sampler, Loras, embeddings, etc.

Incidentally, I believe that Foocus tries to use a smaller LLM To try to expand a prompt.

The problem is not knowing all the keywords, otherwise I'd just use them all at the start. That's the point of having something that guides you.

Keywords is a little too handwavy for my taste - there's no universal definition of what this means.

For example, a lot of the stable diffusion 1.5 checkpoints can generate better looking imagery if you add in essentially keywords such as "trending, masterpiece, award-winning, etc.", but they aren't nearly as important in certain SDXL checkpoints.

To me it sounds like you want an interactive wizard / tutor which could be useful, but my point is you'd have to customize or fine-tune it at a very deep level per model (in the case of images generation). It would definitely be a non-trivial amount of work.

That's why ChatGPT DALL-E 3 is so nice, it's end-to-end in terms of allowing simple expression of what you want to see and an iterative workflow. My younger siblings are in elementary school and are able to get great results from it.

Midjourney also lets you upload an unrelated picture as a stylistic reference, so you can find a picture that has the aesthetic you want (ink, watercolor, matte, etc.)

There are a few, here's one: https://github.com/imrayya/stable-diffusion-webui-Prompt_Gen...

Those prompt suggestions add in a bunch of parameters that you have no idea what they do though, like I'm sure you have no idea who those artists are or what their style looks like, what effects those camera models will have, or what zenethic rays are

The people who might pay for such a tool would be heavy users, who probably need something more than basic advice.

Building something that can actually help those users seems like it would be challenging.

This is also good for minimal prompts - https://mitenmit.github.io/gpt/

Just iterate with the model you want to use. There's no need to add another layer of models to interact with a model.

https://poe.com/PromptGenius

You can do that with ChatGPT just prompting. Whatever tool you are looking for will be just RAG using ChatGPT.

Reply on news.ycombinator.com