Join our Discord community for any help:
https://discord.gg/RnUHPzTkXH
How to install?
Download and extract the .zip, you will find 2 folders, one has ComfyUI custom nodes, and the other one contains the workflow. Copy the custom nodes folder to your ComfyUI/custom_nodes directory, and install their requirements by doing pip install -r requirements.txt. After that, open the workflow in ComfyUI and start using the node.
🧠Ideogram 4 Autoprompter custom node and workflow — How It Works
This node supercharges your Ideogram 4 prompting by auto-generating
a structured, region-aware JSON prompt from a simple idea or image.
Two ways to run it:
- Text only → Type a brief concept or instruction (e.g. "asian girl, doll like face")
and the AI expands it into a full scene with regions, style, and lighting.
- Image + Text → Upload a reference image and the node captions it,
then builds a prompt that matches and enhances what it sees.
Engine options:
- Local → Uses a local HuggingFace vision model (auto-downloaded on
first run). Free, private, no API key needed.
- Gemini → Uses Google's Gemini API for higher quality outputs.
Get a free key at aistudio.google.com — Gemini 3.5 Flash recommended.
After generation, nothing is locked.
You can freely edit the bounding boxes, region descriptions, style tags, lighting, and colors before hitting Generate. The AI gives you a strong starting point — you stay in control of the final result.
Description
FAQ
Comments (30)
where do you get the IdeogramDualModelGuider node it dosent exist anywhere? and installign requirements doesnt fix it
When you download the .zip file from here in civitai, you will see a folder named custom nodes, inside that there are 2 folders which are comfyui-ideogram-autoprompter and comfyui-ideogram-guidance you need to copy both of these folders to your comfyui custom nodes folder and install both the requirements. After that it will work. I will be getting these nodes listed in comfyui manager very soon. I honestly don't know the process. I'll do some research and get it listed. For now try this solution.
@PastellPastellPastellI My guess is that you are using the desktop or an older version of ComfyUI. I recommend using the portable version and/or update to the latest portable version if needed.
#justpassingthrough
@coolstrad thank you
I tested it and it works perfectly, just use the workflow and install the nodes that come with the zipped file.
I'm using the local model and it's working great, generating the entire prompt from the image with all the bboxes correctly.
What local model are you using, GEMMA 12B?
@zerocool22Â qwen3vl 4b abliterated, But I will also test it with other models to see which one performs best.But an abliterated model is what I actually need.
LOVE THIS NODE.
Quick question, any suggestions on fixing the output of the prompter when it converts a reference image sideways? It has the same aspect ration as the output. But still generating the new image sideways.
same here
@RKT_BLUE @JBsharp I recommend you guys try out the Gemma4:26b the issue you guys mentioned can happen with using Low parameter models. Using a big model will solve the issue. Also, a recommendation use Ollama for local models, the inference with it is a lot faster and efficient.
@coolstrad You are crazy, GEMMA 26B, you think all of us have a 5090 GPU?
@zerocool22Â Gemma4 26B Q4 QAT (14GB) with 16k context easily in RTX4080 16GB Vram. Just 15% inteligence lost.
@zerocool22Â Gemma4 26B needs a 16GB GPU, you dont need a 5090 to run that.
@coolstrad bro, did you now how MUCH time are expend in a 26b model ? NOT WORTH
I think 2 neat features to add would be:
- Adding a text input to the node. User can automate wildcards and send random texts to the node every run.
- Adding a dedicated nsfw secondary processing node that takes 1) sfw json generated by Gemini free api and 2) user nsfw edit, process both to give a final json with user intent in correct json format.
Hi, first of all, thank you for creating this workflow and custom node. I've been experimenting with it a lot recently, and it's one of the most interesting Ideogram workflows I've seen so far.
I have a feature suggestion that might make the workflow even more powerful.
Currently, the autoprompter generates regions using bounding boxes (BBox), which works well for basic object placement. However, for more complex scenes, characters, manga panels, and detailed composition control, BBox regions can be quite limiting.
Would it be possible to add support for the region tools used in Ideogram 4 Prompt Builder KJ V2? (from github.com/kijai/ComfyUI-KJNodes/issues/667 this.)
For example:
Polygon regions
Circle regions
Brush-painted regions
Freeform mask regions
Editable region shapes instead of only rectangular BBoxes
It would also be amazing if the autoprompter could automatically generate a dedicated background region in addition to object regions.
Something like:
Character regions
Object regions
Background region
Environment region
This could allow users to automatically analyze an image, generate region data, and then fine-tune the result using the more advanced editing tools from Prompt Builder KJ V2.
In my opinion, combining automatic image analysis from the Autoprompter with the flexible region editing tools from Prompt Builder KJ V2 would create an extremely powerful workflow for image recreation, manga reconstruction, and scene-based Ideogram prompting.
Just a suggestion from someone who has been using both nodes extensively. Thanks again for your work!
You're welcome, I'll try to implement these ideas on the next update. I also thank you a lot for sharing all of these, it helps me improve the work a lot more.
How to avoid safety filter after editing images?
Adding objects on the canvas will automatically bypass that issue
Please add support for llama.cpp server. Ollama is not very popular nowadays. Thx in advance.
Absolutely. Will roll it out on the next update
@coolstrad yes please, llama.cpp server is good
Sorry for the noob question, how do we use loras with this workflow?
Use load lora nodes
after your unet / Diffusion models nodes and plug it to the next models input
Is there a way to use this with LM Studio instead of Ollama?
+1 on that
Absolute gamechanger! Thank you.
Would you consider making it capable of using Grok API for prompting since grok is uncensored? I love this workflow by the way, ideogram was frustrating me until this.
this, gemini not even let us type big breasts
Can you fix local LLM implementation?
It always checks to download from HF first, even after it's downloaded, so you can't generate when offline.
It also never copies the downloaded LLM to a permanent location (portable ComfyUI). It's stuck in users\windowsaccount\.cache on Windows.
There also should be option to point to existing LLM folders in ComfyUI. Shouldn't need to download LLM's we already have, and downloading it via the node is much slower and took hours.