I’m trying to build a workflow to generate images for blog articles but I encounter multiple issues:
Images are generated with AI often include misleading information. AI often hallucinates and messes up letters or invents numbers that are not in the article.
Prompts that are generated by AI inside the article writing flow are often too broad, which makes images not related to the topic. I tried fixing prompts but it’s just to difficult. Maybe there’s a way to do this without prompts? Like a style repository? But again, images must be unique for each article and AI must know what to have on them besides colors and shades.
Images are not being cropped.
Images are too heavy for publishing which affects SEO.
@flowRunner ditch AI image gen for blog posts honestly, the text/number hallucination problem doesn’t have a clean fix yet. use the HTTP Request node to pull from Unsplash API based on your article keywords — images are royalty free, no hallucinated text, and you can resize + compress in the same workflow with an Edit Image node. way more reliable than fighting prompt engineering for every post.
Hi @flowRunner i think most of this can be solved by good prompting practices, but for the too heavy images you can try downgrading to a lower model, but that would decrease the quality but the size problem would be solved, you can try different aspect ratios, but that problem would be there as long as you want images on the article (good quality)
@achamm I like your way of thinking but it does not solve the problem because inline images should have some numbers or text related to the article. I can use QuickChart API for some clients to where I can pass clean data in JSON with numbers from the article but this is just a partial solution for some use cases.
I tried prompting for some years now but it’s unreliable. I need to automate it so well that zero human review is required. With prompts, it’s a 50/50 gamble with each generation.
@flowRunner ok so for the text/numbers problem, QuickChart is actually the right call — you can have a Code node extract key figures from the article, build the chart JSON, then hit QuickChart via HTTP Request. for the non-chart images, layer the quickchart output on top of your Unsplash pulls using Edit Image node composite mode. that way numbers are always clean data, never AI-generated.
swap out the regex for whatever data extraction fits your articles, and the resize handles the file size/SEO issue too.
QuickChart should already generate images (pie charts, bar charts, etc.) so how I would add the “quickchart output on top of your Unsplash pulls”? I can use JSON maybe for that? But maybe not all the numbers in that JSON like I would use for QuickChart because QuickChart would require multiple datasets for axes in order to generate something solid.
@flowRunner right, QuickChart gives you a PNG back — set backgroundColor to transparent in your chart JSON config, then use Edit Image node with the Composite operation to place that transparent chart PNG on top of the Unsplash photo. two separate HTTP Request nodes (one for Unsplash, one for QuickChart), both feed into Edit Image. no extra JSON layering needed, the node handles it natively.