> ## Documentation Index > Fetch the complete documentation index at: https://comfyui-mcp.artokun.io/docs/llms.txt > Use this file to discover all available pages before exploring further. # How to Run Ideogram 4 Locally in ComfyUI (2026 Guide) > Run Ideogram 4 — the 9.3B open-weight text-to-image model — locally in ComfyUI. Install, VRAM tips, JSON area prompting, and flawless readable text. *by [artokun](https://github.com/artokun) · June 16, 2026 · ideogram · text-to-image · ComfyUI · model highlight* Most image models can paint a gorgeous scene and then butcher the one word you actually needed on the sign. **Ideogram 4** is the open-weight model that finally doesn't — and you can run the whole thing **locally in ComfyUI**, no API key, no per-image fee. This is the first entry in our model-highlight series, and a fitting flagship: Ideogram 4 is, by the open-weight leaderboards, the best you can self-host today for **text-in-image and graphic-design layout**. Below: what it's genuinely best at (with receipts), how it stacks up against Flux / Qwen / Z-Image / ERNIE, the VRAM you need, and the fastest way to get it running — a one-command install with [comfyui-mcp](https://github.com/artokun/comfyui-mcp) and the [sidebar Panel](../panel), instead of hand-downloading \~50 GB of weights and hand-typing JSON coordinates. > **TL;DR — one-command setup.** Install comfyui-mcp, apply the `ideogram` pack > (`apply_manifest --path packs/ideogram/manifest.yaml`, or run the generated > `install-windows.bat` / `install-runpod.sh`), and drive the graph from your own > Claude session via the Panel. Jump to [Install](#install-ideogram-4-in-comfyui). ## What is Ideogram 4? Ideogram 4 (released June 3, 2026) is the first **open-weight** foundation model from Ideogram — a **9.3B-parameter single-stream Diffusion Transformer**, trained from scratch, with a **Qwen3-VL-8B** text encoder. It renders at 256–2048px and, uniquely, reads a **structured JSON prompt** that lets you place text, objects, and colors in specific regions of the frame. **Open weight, not "open source" — read this first.** Ideogram 4 ships under a custom **Non-Commercial Model Agreement**, *not* Apache/MIT. Personal use, research, and evaluation are fine; **commercial use needs a separate paid license**, and the terms forbid using its outputs to train competing models. The HuggingFace download is gated. There's also an **embedded safety filter you can't disable from ComfyUI**. None of that changes how good it is — but if your use case is commercial or you expect an uncensored local model, know it going in. ### Why text and layout are its superpower This is the headline, and it holds up to scrutiny: * **\~0.97 English OCR accuracy** (X-Omni) — the best of any open-weight model at its scale, ahead of much larger ones (Qwen-Image 20B, FLUX.2 dev 32B, HunyuanImage 3.0 80B MoE). * In a designer-preference typography study, it was picked **first \~48% of the time** and scored highest on "would you use this in real client work." * **#1 open-weight model on Design Arena** (Elo ≈ 1285, a \~115-point lead over the next open model). A fair caveat: the typography study is vendor-run and Design Arena is design-weighted, so the "beats Flux/Qwen" framing is strongest **for text and layout specifically** — not general photorealism. Reviewers consistently scope the win that way, and so do we. ## Ideogram 4 vs Flux, Qwen-Image, Z-Image, and ERNIE There's no single "best" — pick by the job: | Pick… | When you need… | | ----------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Ideogram 4** | Readable in-image **text**, posters/logos/packaging, **layout & spatial control** (bounding boxes), brand-color accuracy, multilingual headlines, 2K print output | | **FLUX.2** | **Photorealism** and a mature production default | | **Qwen-Image** | Text + multilingual on a larger, popular dev model | | **Z-Image Turbo** | **Speed + low VRAM** (\~2.5s, \~8 GB) for fast ideation | | **Z-Image Base** | Predictable control and **LoRA-training** stability | | **ERNIE-Image** | **Truly permissive (Apache 2.0)** commercial use + CN/JP/EN text | Where rivals win: photoreal portraits (Ideogram trails), raw speed and low VRAM (Z-Image), permissive licensing (ERNIE, Z-Image), and editability — Ideogram's output is flat pixels, not layers. Each of those models gets its own post in this series; Ideogram 4 earns the spotlight for **information-carrying images**. ## System & VRAM requirements The full FP8 release is two \~13.8 GB diffusion models plus the encoder — **\~50 GB on disk**. VRAM: | Build | VRAM | Notes | | ----------------- | ----------------------------------------- | ------------------------------------------------------------- | | FP8 (official) | **16 GB floor**, **24–32 GB comfortable** | Two \~13.8 GB UNETs + 8 GB encoder; ComfyUI swaps them in/out | | NF4 | \~12 GB | Quality/VRAM trade | | GGUF (e.g. Q4\_K) | **\~8 GB** | Community quants for smaller cards | Treat 16 GB as the minimum and 24 GB+ as the smooth-at-2K target. On a 24 GB RTX 4090 you're in good shape. ## Install Ideogram 4 in ComfyUI The manual route works: update ComfyUI, download the five files into the right folders, and load the text-to-image workflow. | File | Folder | | ------------------------------------------------------------------------ | -------------------------- | | `ideogram4_fp8_scaled.safetensors` (conditional UNet) | `models/diffusion_models/` | | `ideogram4_unconditional_fp8_scaled.safetensors` (unconditional UNet) | `models/diffusion_models/` | | `qwen3vl_8b_fp8_scaled.safetensors` (text encoder) | `models/text_encoders/` | | `gemma4_e4b_it_fp8_scaled.safetensors` (prompt-builder helper, optional) | `models/text_encoders/` | | `flux2-vae.safetensors` | `models/vae/` | (Canonical weights live in the **Comfy-Org** HF repos.) ### The fast way — comfyui-mcp + the Panel Downloading 50 GB to the right folders and wiring two UNETs by hand is exactly the busywork the [comfyui-mcp](https://github.com/artokun/comfyui-mcp) **`ideogram` pack** removes. One declarative manifest installs the custom nodes (incl. the KJ Prompt Builder) and pulls every model to the correct folder — and the same manifest drives both an MCP-native install and the generated double-click scripts: ```bash theme={null} # MCP-native (from a Claude Code session, with COMFYUI_PATH set) apply_manifest --path packs/ideogram/manifest.yaml # or one-click packs/ideogram/install-windows.bat # Windows packs/ideogram/install-runpod.sh # RunPod / Linux ``` Then load `packs/ideogram/workflow.json`. Because the pack ships with the [plugin](../plugin), your **own Claude session can drive the live graph through the [Panel](../panel)** — add/wire nodes, set widgets, and iterate on prompts conversationally, with full Ctrl+Z undo and no extra API keys. Every model URL in the pack is CI-validated for reachability and size, so a link never quietly rots ([here's why that matters](./installer-packs-that-cant-rot)). ## Structured JSON prompting (the part that sets Ideogram apart) Ideogram 4 was trained on **structured JSON captions**, so instead of one loose sentence you hand it a document that "deconstructs" the image into regions — each described, placed, and colored separately. That's what makes text and layout controllable: ```json theme={null} { "high_level_description": "one-line summary of the whole image", "style_description": { "aesthetics": "...", "lighting": "...", "art_style": "...", "color_palette": ["#0A1F44", "#FFD200"] }, "compositional_deconstruction": { "background": "scene, palette, lighting, mood (required)", "elements": [ { "type": "obj", "bbox": [200, 120, 760, 880], "desc": "a red sports car, 3/4 view" }, { "type": "text", "bbox": [40, 100, 180, 900], "text": "GRAND OPENING", "desc": "bold condensed sans, gold", "color_palette": ["#FFD200"] } ] } } ``` Two things to get right: * **`bbox = [top, left, bottom, right]`**, on a **0–1000** normalized canvas (not pixels, not x/y/w/h). Canvas width/height must be multiples of 16. * **Palettes** condition color: up to 16 hex at the image level, up to 5 per element. You don't have to type coordinates by hand. The **Ideogram 4 Prompt Builder (KJ)** node — bundled in the pack — is a visual canvas: in V2 you draw regions with brush/rectangle/ellipse/polyline, get per-region layers (overlaps don't merge), bucket fill, an eyedropper, and an opacity slider to trace a reference image. Each shape becomes a bbox automatically. The pack also includes **25 ready-made JSON templates** (drop them in `ComfyUI/user/default/kjnodes/ideogram4/templates`). **Area-prompting tips that actually help:** * One bbox per *major* subject — don't split a person into face/hair/clothes. * Reserve zones for posters: title up top, subject center, CTA at the bottom. * 3–6 strong, non-overlapping elements beat 20 overlapping ones (overlap garbles text). * Lock global look with `style_description` so per-element prompts don't drift. ## What to make with it The pack's 25 templates map to its sweet spots: * **Logos & brand identity** — wordmarks, mascot kits, luxury identity sheets * **Posters & key visuals** — film/event posters with legible headlines * **Covers** — book, magazine, comic, anime covers * **Product & display ads** — packaging, e-commerce creatives, menus * **Character & game sheets** — 2D/3D character sheets, trading cards, item sheets * **Merch & signage** — t-shirt graphics, print-on-demand, thumbnails If the image has to *carry information* — a headline, a price, a brand — this is the open model to reach for. ## Settings that matter The pack's workflow is tuned already, but for reference: sampler **euler**, scheduler **simple**, **28 steps** (turbo: 12), **cfg 5**, `ModelSamplingAuraFlow` shift **5**, denoise 1.0 (0.6 for img2img). Ideogram 4 uses **two UNETs** — the unconditional model feeds a `DualModelGuider` for the CFG baseline, so there's **no negative text prompt** (it's zeroed out). The V2 workflow applies any LoRA to **both** the conditional and unconditional models. ## Troubleshooting * **Text has a typo / is garbled.** Keep rendered strings short and bold; don't overlap text boxes; proofread before publishing — high OCR ≠ perfect every time. * **Gated download / 401.** Accept the license on the HuggingFace page and authenticate, or let the pack pull from its mirror. * **Safety filter refusals.** Baked into the weights; it can't be disabled from ComfyUI. * **OOM at 2K.** Drop to a GGUF/NF4 build, generate at lower res first, or use a 24 GB+ GPU. * **Same seed, different image across machines.** Expected — output varies with GPU/driver/CUDA/ComfyUI version. ## FAQ **Is Ideogram 4 open source or open weight?** Open *weight*. The weights are downloadable under a custom **Non-Commercial** license — not an OSI open-source license. **Is it free to use?** Free to download and use non-commercially. Commercial use requires a separate paid agreement with Ideogram. **Can I run it locally / offline?** Yes — fully local in ComfyUI, no API key and no network call at generation time. **How much VRAM do I need?** \~16 GB minimum for the FP8 build, 24 GB+ for smooth 2K work; GGUF quants bring it down to \~8 GB. **How many parameters?** 9.3B — a single-stream Diffusion Transformer with a Qwen3-VL-8B text encoder. **Is it better than Flux or Qwen-Image?** For **in-image text and design layout**, yes by the open-weight leaderboards. For photorealism, Flux is the stronger default. **Why is its text so accurate?** It was trained on structured JSON captions and optimized for typography, hitting \~0.97 OCR accuracy — text generation is the model's core design goal, not an afterthought. *** ## Get it running in one command 1. Install [comfyui-mcp](https://github.com/artokun/comfyui-mcp) and the [Panel](../panel) — the panel auto-starts a background agent on your Claude subscription (no API keys; sign in with `claude` once). 2. Apply the **`ideogram` pack** — nodes + \~50 GB of models land in the right folders, validated. 3. Open the [Panel](../panel) and let the panel's agent build area-prompts and wire the graph for you. That's the whole point of the project: expert ComfyUI setups that install in one step and drive themselves from your own agent session. **That wraps the model-highlight series** — nine local ComfyUI setups, each one command away. Browse them all on the [blog index](./), or start from the top with [WAN 2.2 in ComfyUI](./wan-2.2-comfyui).