Closed joshwallerr closed 1 year ago
For now, I'll add an extra stage in to generate a unique image prompt, for example:
write me a short, one sentence long text-to-image prompt that best visualises the following headline: {headline}
This seems to get better results than using a template and leaving the creativity up to the model.
I've also changed the style from "an oil painting" to "64-bit pixel art", as this seems to fit the theme of the site better and looks slightly more appealing.
Twitter user complained that the pixel art comes across as insensitive (https://twitter.com/devsysape/status/1658536380778287104) when the article is about war or other sensitive topics.
I will test a few other image styles and look for any that are more suitable.
Also commit fffd1ce upgraded the image gen engine to SDXL beta and custom prompts are now used, so oil paintings shouldn't come out as creepy anymore.
Let's face it, the article images are haunting.
The current prompt for generating images is:
We're using Stable Diffusion via the stability-sdk PyPI package.
Please leave a comment with any other image engines or prompts that produce a more suitable output.