<h1 id="ideogram-4-on-apple-silicon-(mac)-regional-bbox-+-local-llm-workflow">Ideogram 4 on Apple Silicon (Mac) — Regional bbox + Local LLM Workflow</h1>Run Ideogram 4 locally on a Mac (M-series, Apple Silicon) in ComfyUI. Draw boxes to place people and objects, type a short label in each, and a local uncensored LLM expands them into Ideogram's structured JSON before Ideogram 4 (GGUF) renders the scene. No NVIDIA GPU required.Other Ideogram 4 workflows assume an NVIDIA GPU, or the stock GGUF loader which fails on Apple's MPS backend. This one is built for Mac and ships the patches that make it work.<h2 id="what-it-does">What it does</h2><pre><code>1) BBOX EDITOR draw a box per person/object, type a SHORT label in each
 | (e.g. "redhead woman", "wine glass", "small dog")
 v
2) GEMMA a local uncensored LLM expands every region + the whole image
 | into rich JSON, keeping each box exactly where you placed it
 v
3) IDEOGRAM 4 renders each element in its region</code></pre><ul><li>Regional composition: position multiple subjects/objects precisely (box coordinates are preserved verbatim).</li><li>No JSON by hand: Ideogram 4 ignores plain text (it returns a grey "safety filter" box) and only behaves with a structured JSON prompt. The LLM writes that JSON for you from short labels.</li><li>Realistic, non-"AI-looking" results, plus Ideogram's strong text rendering (use a <code>text</code> box to render a sign / signature).</li><li>SFW by default. NSFW-capable (see below).</li><li>The architecture patch is platform-agnostic, so it also works on CUDA — but the focus is Apple Silicon.</li></ul><h2 id="sfw-nsfw">SFW / NSFW</h2>Out of the box, neutral labels give safe-for-work images. For NSFW: the local Gemma is uncensored, and the optional Vintage Beauties / mi55ionary LoRAs (on CivitAI) add explicit anatomy. Load one on the LoRA node and use explicit labels.<h2 id="requirements">Requirements</h2><ul><li>Mac with Apple Silicon (M1/M2/M3/M4/M5), 32 GB unified memory or more recommended.</li><li>ComfyUI (manual install or the Desktop app).</li><li>~30 GB free disk for the models.</li></ul><h2 id="install-easy-(script)">Install — easy (script)</h2>Download the attached package, then from a terminal:<pre><code>cd ideogram4-mac-workflow
chmod +x install_mac.sh
./install_mac.sh # auto-detects ComfyUI, or:
./install_mac.sh /path/to/ComfyUI</code></pre>It installs the three custom nodes, downloads every model to the right folder, and drops in the workflow + LLM prompts. Then restart ComfyUI and load <code>Ideogram4_GGUF_Mac_Bbox_Gemma</code>. (Optional NSFW LoRAs: <code>export CIVITAI_TOKEN=...</code> before running.)<h2 id="install-manual">Install — manual</h2>1. Custom nodes (into <code>ComfyUI/custom_nodes</code>):<pre><code>git clone https://github.com/fxd0h/ComfyUI-GGUF # Ideogram-4 GGUF + an MPS dtype fix
git clone https://github.com/fxd0h/ComfyUI-LLM-text-processor # runs the Gemma GGUF on Apple Silicon
git clone https://github.com/kijai/ComfyUI-KJNodes # the bbox prompt builder</code></pre>The first two are patched forks; the patches are also submitted upstream (city96/ComfyUI-GGUF PRs #455 and #456).2. Models (download and place):<ul><li><code>ideogram4-iQ4_NL.gguf</code> + <code>ideogram4_unconditional-iQ4_NL.gguf</code> → <code>models/unet/ideogram4-gguf/</code> — from <code>stduhpf/ideogram-4-gguf</code> (Hugging Face)</li><li><code>qwen3vl_8b_fp8_scaled.safetensors</code> → <code>models/text_encoders/</code> — from <code>Comfy-Org/Ideogram-4</code> (path <code>text_encoders/</code>)</li><li><code>flux2-vae.safetensors</code> → <code>models/vae/</code> — from <code>Comfy-Org/Ideogram-4</code> (path <code>vae/</code>)</li><li><code>gemma-4-E4B-it-ultra-uncensored-heretic-Q4_K_M.gguf</code> → <code>models/LLM/gemma4-e4b-uncensored/</code> — from <code>llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF</code></li><li>LoRAs (optional, NSFW) → <code>models/loras/</code> — from CivitAI (Vintage Beauties, mi55ionary)</li></ul>3. LLM prompts — copy <code>ideogram4_json.txt</code> and <code>ideogram4_expand.txt</code> (in the package, <code>prompts/</code>) into <code>models/LLM/prompts/</code>.4. Workflow — copy <code>Ideogram4_GGUF_Mac_Bbox_Gemma.json</code> into <code>ComfyUI/user/default/workflows/</code>.Restart ComfyUI. The Gemma node auto-downloads a small macOS llama.cpp binary on first run.<h2 id="how-to-use">How to use</h2><ol><li>Bbox editor (node 1): on its canvas — drag to draw a box, double-click to edit (set its label and <code>obj</code> vs <code>text</code>), click to select, Del to remove, Ctrl/Cmd+C/V/D to copy/paste/duplicate. One box per subject/object with a short label.</li><li>Gemma (node 2) expands every region + the whole image. Leave <code>system_prompt = ideogram4_expand.txt</code>.</li><li>Image size / batch: set the SAME width/height in <code>EmptyFlux2LatentImage</code> and <code>Ideogram4Scheduler</code>; batch on the latent.</li><li>CFG / seed / steps: <code>DualModelGuider</code> (~3) / <code>RandomNoise</code> / <code>Ideogram4Scheduler</code>.</li><li>Signature / text: add a <code>text</code> box (e.g. your name) — Ideogram renders it cleanly.</li></ol><h2 id="speed">Speed</h2>Render-tier on Mac: ~7 min per 1024px image at 20 steps (GGUF dequant on MPS is the bottleneck), plus ~10–30 s for the Gemma step. To go faster: swap <code>DualModelGuider</code> for <code>BasicGuider</code> (single model, ~2x), use 12 steps, or lower the resolution. For maximum speed, the <code>mflux</code> CLI runs the same model in ~90 s.<h2 id="troubleshooting">Troubleshooting</h2><ul><li>"Unknown model architecture" on the GGUF node → you're on the stock node; use the patched fork <code>fxd0h/ComfyUI-GGUF</code>.</li><li>Gemma node errors about a binary → it downloads a macOS llama.cpp build on first run; allow it network access.</li><li>Grey image → your prompt reached Ideogram as plain text; keep the Gemma → CLIPTextEncode chain (JSON is what gets through).</li></ul><h2 id="notes">Notes</h2><ul><li>Ideogram 4 has a non-commercial license; respect it.</li><li>Developed with AI assistance and verified on-device on Apple Silicon.</li><li>Forks / upstream: <code>github.com/fxd0h/ComfyUI-GGUF</code> (city96/ComfyUI-GGUF #454/#455/#456), <code>github.com/fxd0h/ComfyUI-LLM-text-processor</code>.</li></ul>

Ideogram 4 on Apple Silicon (Mac) — Regional bbox + Local LLM Workflow

Initial release — Ideogram 4 GGUF + regional bbox + local LLM, for Apple Silicon (Mac).

Files with SHA256: 9faf41dfd565c6cd375a34ac07a08882ee734a41cc56f0c47f29c01ec61ca30f