[enhancement]: Ideogram Model support

[m4iccc]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

This looks like a very promising model, please add support for it! Ty!!!

https://github.com/ideogram-oss/ideogram4
https://huggingface.co/ideogram-ai/ideogram-4-nf4

Alternatives

No response

Additional Content

No response

[m4iccc]

This model already has diffusers support to their fp8 and nf4 models

[doppelrittberger]

Would also love to see the support for this awesome model :)

[sammcj]

Yeah this would be fantastic, super interesting model.

FP8 link: https://huggingface.co/ideogram-ai/ideogram-4-fp8

FYI I am having a crack at a PR to add it, I'll link here if I get it working cleanly.

[sammcj]

I have a mostly working proof of concept running ideogram-4 in Invoke here: main...sammcj:InvokeAI:ideogram4

I was testing using it with Qwen 3.5 4B GGUF to generate the magic prompt syntax which worked OK for testing but obviously for high quality results you'd want to use a larger model.

I also managed to port ideogram4 to to MPS / Apple Silicon here: ideogram-oss/ideogram4#19, it's a bit heavy though due to the amount of f32 tensors, certainly will benefit from proper GGUF quantisation at some point.

I've probably run out of time (and tokens!) for this one now, but hopefully it's a starting point or at least interesting for someone else to pick up and run with.

[KillerX629]

what would a more appropiate solution be for the JSON prompt expansion part? maybe a canvas plugin that will convert square regions to that format?

[m4iccc]

I have a mostly working proof of concept running ideogram-4 in Invoke here: main...sammcj:InvokeAI:ideogram4

I was testing using it with Qwen 3.5 4B GGUF to generate the magic prompt syntax which worked OK for testing but obviously for high quality results you'd want to use a larger model.

I also managed to port ideogram4 to to MPS / Apple Silicon here: [ideogram-oss/ideogram4#19](https://github.com/ideogram-oss/ideogram4/pull/19), it's a bit heavy though due to the amount of f32 tensors, certainly will benefit from proper GGUF quantisation at some point.

I've probably run out of time (and tokens!) for this one now, but hopefully it's a starting point or at least interesting for someone else to pick up and run with.

That's pretty cool, I think the two must have options to port over that model is the JSON translator and the Region prompt drawer that KJ made specially for this model: