UPDATE COMFYUI TO THE LATEST UPDATE!
Ideogram 4 is Ideogram's first open weight text-to-image model. It is a state-of-the-art foundation model trained from scratch — not a fine-tune of any existing model. It introduces a new structured JSON prompting interface, with best-in-class multilingual text rendering, deep language understanding, explicit bounding-box layout and color-palette controls, and native 2k resolution images. The easiest way to try the model is online at ideogram.ai.
Diffusion Model (nvfp4) - REQUIRED!
Unconditional Diffusion Model (nvfp4) - REQUIRED!
https://huggingface.co/Comfy-Org/Ideogram-4/blob/main/diffusion_models/ideogram4_unconditional_nvfp4_mixed.safetensors
Text Encoder (fp8)
https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/blob/main/full_encoder_small_decoder.safetensors
requires both diffusion models. youll understand why when you see the workflow deconstructed. ive set this workflow up as i dont like the subgraph workflows. takes away from the joy of comfy imo <3
Description
gguf coming shortly
FAQ
Comments (10)
I'm using the workflows provided on comfy site since yesterday, had to add a clean ram node to prevent "buffer exhausted" error or something like that. But my images are atrocious. Crossed eyes, deformed hands, etc. No idea how people are getting such good results. I'll try this one and come back later to say how it goes 🤣
Edit: still looks like dog crap https://civitai.red/posts/29012297
i had to modify the workflow significantly to stop getting content filter blocked
https://civitai.red/posts/29085948
but these images didnt have all the artifactey blurriness once i did that
Update to the latest ComfyUI (the initial workflow template is buggy) and try this: https://www.reddit.com/r/StableDiffusion/comments/1tyo4ha/comment/oq57cs1/?screen_view_count=3
Try using the ModelSamplingAuraFlow node between the model and the CFG and set the shift to 5. Also it doesn't take long at all on Turbo mode (12 steps) and still looks great. I'm running it on 16GB of VRAM and it takes 30 seconds per image.
@jimecherry wtf, my computer is trolling me. Edit: I guess I finally got it. I was dumb and trying to use the nvfp4 model but I don't have a Blackwell gpu. It's perfect now with the fp8 model. Thanks for trying the same prompt and workflow.
@NowhereManGo sounds great. I'm on 8GB Vram and it's taking 500 seconds...
@WetnWild nv4p stills runs on non blackwell i'm on amd so its either nv4p or a gguf when those come out. takes 4mins at 1.5megapixels
@jimecherry Yes, just that it can be slow without native nv4p support (which is only on 500x, IIRC).
Hello, here my experiences with the workflow. It runs and needs 265 seconds to render the clock with the mirrored city (MSI notebook, 4080 with 12 GB VRAM, 32 GB RAM). But i get an error message if i want to render the prompt from jimecherry: "Image blocked by safety filter" like he says it in his comment. The workflow from jimecherry runs with this prompt. In my opinion the results are not better than f.e. ZIT and needs more time. And Ideogram needs more experience with the text in the prompts. But thanks for your work. Perhaps you can better define the image with enough experience.
Not working at all.
Have no idea what they did with paranoidal filter but i cant even replace clock face by window in example prompt - filter blocks image.
Model seems just completely useless it literally cant even alter example prompts. Nothing except this roofs behind clock can be generated.
Thank you for putting this together - I am ok until it gets to the sampler and then vram crashes by about 700MB (I have 4GB into total) - I manage to run Flux GGUF and ZIT - also using tiled Vae decode. I hope GGUF version of Ideogram works as it looks like a great model. Thanks again.