8-15 step T2I with neat quality
✅ Just download all the models https://huggingface.co/Comfy-Org/Ideogram-4 (you can skip ideogram4_unconditional_fp8_scaled.safetensors for v1.0/1.3-speed)
✅ Update ComfyUI
✅ v1.3: Update Sage Attention
✅ Install/Update KJ Nodes (https://github.com/kijai/ComfyUI-KJNodes)
✅ Install/Update AGSoft Nodes (https://github.com/Art-xmaster/comfyui-AGSoft)
✅ v1.3: Install/Update RaykoStudio Nodes (https://github.com/Raykosan/ComfyUI_RaykoStudio)
✅ Install (for v1.0/v1.1) https://github.com/giriss/comfy-image-saver or (for v1.2 and higher) https://github.com/alexopus/ComfyUI-Image-Saver
1.3-speed:
Very fast gen, slightly faster than Flux.1, 3x faster than base speed
Just update Sage Attention and download required LoRAs (links inside the workflow)
📜 Prompting Guide: https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md
Description
Patch Flash Attention KJ Node - for faster gens
Two Load Lora nodes (for both main models)
comfy-image-saver nodes (deprecated) changes to ComfyUI-Image-Saver
Manual Prompt node is now connected to the KJ Prompt Builder (please read the note inside WF near Ideogram 4 Editor)
Prompt nodes rearranged a little
Added Save Image node to save the result
Links and instructions inside the WF
Proper width/height detection
FAQ
Comments (5)
Where to find ideogram__qwen3vl_8b_fp8_scaled.safetensors?
Also, there is an "'NoneType' object is not callable" error in Patch Flash Attention KJ. Would appreciate an advise.
TE you can find here https://huggingface.co/Comfy-Org/Ideogram-4/blob/main/text_encoders/qwen3vl_8b_fp8_scaled.safetensors (ideogram__qwen3vl_8b_fp8_scaled.safetensors is just a naming in my folder to understand which model this TE belongs to)
'NoneType' object is not callable - you should update KJ Nodes
"SamplerCustomAdvanced is missing a required input: latent_image"
As I go with "Fix node (recreate)" on the SCA node, it reaveals unlinked socket with latent_image. What can I do? Please advise :(
Maybe ComfyUI version missmatch? I updated it today
I think you missed cg-use-everywhere nodes