the pormpt "Photo of woman lying om grass" gets blocked. i guess easily outperforms Flux2, SD3, Zimage
You can clearly see Ideogram outperforming sd3 in terms of safety, alignment and responsibility. I am so glad Ideogram team protects us from generating indecent imagery such as "w*men"! What would children think if they saw this horrible, lewd image the other model produced!?
just need to structure the prompt. I also had this filter thing pop up a bit.
But after structuring the prompt better (or use LLM for a json prompt), i never see this anymore ;-)
https://huggingface.co/ideogram-ai/ideogram-4-fp8#prompting-guide
https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md
You can try this system prompt with any LLM node in comfy, or using any llm site:
https://pastebin.com/VihLY5Ct
Ideogram truly seems to support and push for gender discrimination with this move.
i guess this model is just TOO optimized for json prompt and have lost the ability to classify natural language prompts, anyhow, a built-in safety filter is a scum move.
btw, does any of what ideogram did rings any bells? isnt this what stable diffusion 3 did all over again but worse?
- non commercial license
- built in safity filter
- no bf16 weights
- json ONLY
Its json prompting "only"
And it creates stunning images if just prompting right
https://blog.comfy.org/p/ideogram-4-day-0-support-in-comfyui
or try here https://huggingface.co/spaces/ideogram-ai/ideogram4
I'm never seeing the filter ever, when I made the prompt json with LLM node in comfy.
Its probably not a safety filter in the way it sounds, the model just fails when the prompt structure is unknown / not correct
But they are working on fixing that
The safety filter error was a bit unfortunate, since it might give the wrong first impression
But its easily one of the best open source models, and with regional prompting, its really something
The json prompting is mandatory indeed, most of the "safety" filter issues are just the model collapsing on out of distribution prompts. Quick example with the json structure using my prompt builder:
Not suggesting my settings are optimal at all, and this is relatively lazy attempt, just to show it's less a safety issue than the model just not working with plain/short prompting.
I’ve observed the same thing. With proper JSON formatting and a reasonable request, there can still be some issues with the safety filters, but even so, I can easily see that this is one of the best open-source models for consistent text rendering and creating stunning images.
That said, it’s a bit unfortunate that the first impression may not be as impressive as it could be. Still, I’m genuinely grateful that the team chose to release their latest models, which they paid tremendous amount of their resources.
Thank you very much for this amazing work!