Live data from Hacker News

Show HN: Ideogram 4.0 – open-weight 9.3B text-to-image model

github.com

11–15 of 15 posts

Re: Show HN: Ideogram 4.0 – open-weight 9.3B text-to-image model

#13
post #8
post #7

Nice to see another locally hostable model! It’s going to take me a bit longer to add this model to the GenAI Showdown benchmark [1], since I’ll need to add a bit of customization so it produces highly optimized JSON-structured prompts. It might be worth noting that fal.ai [2] ( a fairly popular router in the generative AI space ) doesn’t really mention or emphasize the JSON-structured prompt format, and seems to sug…

[1] Nice, let us know if you need a hand You can call our API to generate structured json from text using this endpoint (this endpoint is free of charge, but you need a free account to create an api key): https://developer.ideogram.ai/api-reference/api-reference/ma... [2] Fal runs the text_prompt (natural language) through the magic prompt system, so indeed results should be good there

Sorry for the belated reply!

This is a impressively good model - even more so by virtue that it is locally hostable and is comparable to the original NB released in August of last year. Ideogram doesn’t have the fanfare of Black Forest Labs or some of the other big names, but they consistently produce very solid work.

It took me longer than anticipated to produce a solid pass through the benchmark for a number of reasons. The strong emphasis on proper JSON formatting as opposed to the more common natural language makes a huge difference in output quality.

Another issue is that the default workflow provided by ComfyUI tends to produce heavily “overcooked” images because of the high CFG settings. With some workflow adjustments, though, you can get very strong results.

Of all the recently released locally hostable models, this one definitely demands the most manual effort. But if you’re willing to put in the work, it can produce fantastic results (scoring an 8 out of 15 on the benchmark and the only local model that managed the Bee comic)

https://genai-showdown.specr.net/?models=fd,hd,kd,qi,f2d,zt,...

Re: Show HN: Ideogram 4.0 – open-weight 9.3B text-to-image model

#14
post #8

Earlier quoted context omitted.

[1] Nice, let us know if you need a hand You can call our API to generate structured json from text using this endpoint (this endpoint is free of charge, but you need a free account to create an api key): https://developer.ideogram.ai/api-reference/api-reference/ma... [2] Fal runs the text_prompt (natural language) through the magic prompt system, so indeed results should be good there

Sorry for the belated reply! This is a impressively good model - even more so by virtue that it is locally hostable and is comparable to the original NB released in August of last year. Ideogram doesn’t have the fanfare of Black Forest Labs or some of the other big names, but they consistently produce very solid work. It took me longer than anticipated to produce a solid pass through the benchmark for a number of rea…

Thank you for doing this <3 Love this showdown
Post reply on HN