ComfyUI LlamaCPP → Ideogram Prompt

A ComfyUI custom node that calls your local llama.cpp server to expand a short idea into a rich, structured JSON prompt for Ideogram 4 — with full thinking/reasoning support.

What it does

• Connects to your local llama.cpp server (/v1/chat/completions)
• Selects from live models via dropdown (auto-populated from /v1/models)
• Enables deep reasoning via thinking_budget_tokens — Qwen3, DeepSeek-R1 and similar models will actually use their <think> block
• Strips <think>...</think> blocks so only clean JSON reaches Ideogram
• Validates and minifies the JSON output
• Optionally unloads the model after generation to free VRAM before diffusion runs

Requirements

• ComfyUI
• llama.cpp server running locally in router mode
• A thinking-capable model loaded (tested with unsloth/Qwen3.6-27B-MTP-GGUF, unsloth/Qwen3.6-35B-A3B-GGUF)
• Ideogram 4 set up in ComfyUI (model + VAE + CLIP)

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/iChristGit/comfyui-llamacpp-ideogram

Restart ComfyUI. The node appears under LlamaCPP / Ideogram.

Node: 🦙 LlamaCPP → Ideogram Prompt

Inputs

Input	Type	Default	Description
user_idea	STRING	—	Short natural-language description of what you want
aspect_ratio	STRING	9:16	Target aspect ratio (e.g. 1:1, 16:9, 9:16, 4:5)
model	DROPDOWN	—	Live model list from your llama.cpp server
server_url	STRING	http://127.0.0.1:8082	Base URL of your llama.cpp server
temperature	FLOAT	0.6	Sampling temperature
max_tokens	INT	8192	Max tokens including thinking tokens
unload_after	BOOLEAN	true	Unload model after generation to free VRAM
unload_wait_seconds	INT	3	Seconds to wait after unload for VRAM reclaim
enable_thinking	BOOLEAN	true	Enable deep reasoning (thinking_budget_tokens)
thinking_budget	INT	4096	Max tokens the model may spend reasoning
system_prompt_override	STRING	(optional)	Override the built-in system prompt

Outputs

Output	Type	Description
ideogram_json	STRING	Clean minified JSON ready for Ideogram's CLIPTextEncode
raw_response	STRING	Full raw response including <think> block (for debugging)

llama.cpp server setup

The node expects your llama.cpp server running in router mode on port 8082. Example start-llamacpp.bat:

"C:\Llama CPP\llama-server.exe" --port 8082 --router ...

Thinking budget

The node sends thinking_budget_tokens per request — no server-side --reasoning-budget flag needed. You'll see this in your llama.cpp logs when it's working:

I reasoning-budget: activated, budget=4096 tokens
...
I reasoning-budget: deactivated (natural end)

natural end means the model finished reasoning on its own before hitting the cap — ideal quality. If you see forced end, increase the budget.

Example Workflow

The example_workflows/ideogram4_t2i.json file is a ready-to-use ComfyUI workflow — just drag and drop it into ComfyUI to get started immediately.

It is based on the official ComfyUI Ideogram 4 txt2img workflow, with one addition: the 🦙 LlamaCPP → Ideogram Prompt node wired in as the prompt source. Everything else — the model loader, VAE, CLIP, sampler, and save node — is identical to the stock workflow.

To use it you will need:

• The Ideogram 4 model weights (ideogram4_fp8_scaled.safetensors)
• The unconditional model (ideogram4_unconditional_fp8_scaled.safetensors)
• CLIP model (qwen3vl_8b_fp8_scaled.safetensors)
• VAE (flux2-vae.safetensors)
• Your llama.cpp server running on http://127.0.0.1:8082

Adjust the model filenames to match whatever you have named them locally.

Example output

The images below was generated with:

• Model: unsloth/Qwen3.6-27B-MTP-GGUF:Q4_K_XL
• Thinking budget: 4096
• Prompt: "an episode of SpongeBob/south park was banned in the United States for being quite offensive. Generate a screenshot of that episode."

Tips

• For complex scenes (multi-character, comic panels, text-heavy): increase thinking_budget to 6144–8192
• For simple product shots: thinking_budget of 1024–2048 is plenty and saves time
• Temperature: 0.6 is a good default. Go lower (0.3) for more consistent JSON structure, higher (0.8) for more creative descriptions
• unload_after = true is strongly recommended — the LLM and diffusion model fighting over VRAM will cause OOM errors on most consumer GPUs

License

MIT