Ideogram 4 in ComfyUI: FP8 and NVFP4 Workflow Guide

The Ideogram 4 ComfyUI workflow brings Ideogram's first open-weight image model into a local node-based setup. It supports natural-language prompts, structured JSON prompts, object and text bounding boxes, multiple aspect ratios, and local generation through either FP8 or NVFP4 model files.

Ideogram released the weights and inference code on June 3, 2026. The official model card describes Ideogram 4 as a 9.3-billion-parameter text-to-image foundation model trained from scratch, with native 2K output and explicit controls for layout, color palettes, and in-image text.

License warning: “Open weight” does not mean unrestricted open source. Ideogram 4 uses the Ideogram 4 Non-Commercial license. The license permits specified research, evaluation, personal, charitable, and non-production uses, but it restricts commercial use. Review the complete license before using outputs in monetized creator work, advertising, products, or client projects.

Table of contents

• What this workflow does
• Requirements
• Choose FP8 or NVFP4
• Install the Ideogram 4 ComfyUI workflow
• Build structured prompts and bounding boxes
• Important generation settings
• Troubleshooting
• Limitations and trade-offs
• FAQ

What this workflow does

The supplied workflows expose Ideogram 4's layout system through ComfyUI. Instead of relying only on a long paragraph, you can define the scene at two levels:

1. Global scene description: subject, background, style, aesthetic direction, lighting, and framing.
2. Regional controls: normalized bounding boxes for objects or text, with a description and optional color palette for each region.

The Ideogram4PromptBuilderKJ node from KJNodes converts those controls into the structured JSON format expected by the workflow. The conditional and unconditional Ideogram models then feed a dual-model guider and the sampling chain.

This approach is useful for compositions where position matters: a person in a specific area, a product on a table, a sign at the top of a poster, or text inside a defined region. It improves spatial guidance, but it does not guarantee pixel-perfect placement or identity consistency.

Requirements

Before importing either workflow, prepare:

• an updated ComfyUI installation, preferably a version that includes Ideogram 4 support;
• enough disk space for two diffusion-model files, one text encoder, and the VAE;
• KJNodes for the Ideogram prompt builder and resolution selector;
• RES4LYF for sigma-related utility nodes used by the supplied graphs;
• rgthree-comfy for the seed and label nodes;
• ComfyUI-GGUF only for the supplied NVFP4 workflow's GGUF text encoder; and
• one of the two packaged workflow JSON files.

The official ComfyUI Ideogram 4 tutorial recommends updating ComfyUI before loading the template. If the workflow opens with red or missing nodes, update ComfyUI and the custom-node packages before changing the graph.

Files used by the FP8 workflow

The FP8 workflow actually loads these files:

The current Comfy-Org file listing reports approximately 9.28 GB per FP8 diffusion file. ComfyUI's documentation currently describes larger approximate download sizes, so check the live repository before planning disk space.

The note embedded in the supplied FP8 JSON also lists gemma4_e4b_it_fp8_scaled.safetensors, but that file is not connected by the supplied graph. Do not download it solely for this workflow unless a newer template explicitly uses it.

Files used by the NVFP4 workflow

The RTX 50-series workflow is intended to use:

The Comfy-Org repository lists each NVFP4 diffusion file at approximately 5.49 GB. Smaller weight files can reduce storage and memory pressure, but actual VRAM use and speed still depend on ComfyUI, the GPU, resolution, attention implementation, and model offloading.

Workflow correction required: In the supplied NVFP4 JSON, both UNET loaders currently select ideogram4_nvfp4_mixed.safetensors. Change the unconditional loader to ideogram4_unconditional_nvfp4_mixed.safetensors before generating. The embedded download note also contains stale FP8 filenames, so use the table above and the live Comfy-Org Ideogram 4 repository.

Choose FP8 or NVFP4

The video describes 24 GB cards such as the RTX 3090 and RTX 4090 as suitable for the FP8 setup. Treat that as a demonstration-specific observation, not an official minimum. The official ComfyUI guide lists the files but does not promise a universal 24 GB requirement.

The narration also suggests that NVFP4 is faster and uses less VRAM on RTX 50-series cards. The smaller files support the storage comparison, but no reproducible benchmark was supplied for generation speed or peak memory.

Install the Ideogram 4 ComfyUI workflow

1. Update ComfyUI

Open ComfyUI Manager and update ComfyUI. Do not restart until you have checked the custom nodes, so you can complete the updates in one pass.

2. Install or update the custom nodes

Install or update:

• ComfyUI-KJNodes
• RES4LYF
• rgthree-comfy
• ComfyUI-GGUF for the NVFP4 workflow

Restart ComfyUI after the installs finish.

3. Download the model files

Download the files for your chosen precision. The repackaged ComfyUI files are available from:

• Comfy-Org/Ideogram-4
• Comfy-Org/Qwen3-VL
• Qwen3-VL-8B-Instruct-GGUF for the NVFP4 workflow

The original Ideogram repository is gated: you must accept its conditions and share the requested account information before downloading directly. The repackaged Comfy-Org repository is publicly visible but retains the Ideogram non-commercial license.

4. Import the matching JSON

Drag the appropriate file onto the ComfyUI canvas:

• Ideogram_V4_FP8 (2).json
• Ideogram_V4_NVFP4 RTX 50xx Only.json

Resolve missing-node warnings before adjusting prompts. For the NVFP4 graph, make the unconditional-loader correction described above.

5. Select an aspect ratio and working size

The supplied resolution selector includes portrait, landscape, and square presets. The video recommends beginning around 1 megapixel for iteration and switching to 4 megapixels only after choosing a composition.

That is a practical workflow choice rather than an Ideogram requirement. Larger images need more time and memory, and output quality can vary by aspect ratio.

Build structured prompts and bounding boxes

Official ComfyUI documentation says the local Ideogram 4 workflow accepts both:

• natural language, suitable for simple scenes; and
• structured JSON, suitable for explicit layout, style, color, and text control.

The video incorrectly states that Ideogram 4 only accepts JSON positive prompts. Natural-language prompting is supported, but structured JSON exposes more of the model's layout interface and may reduce false-positive safety blocks.

Fill in the global description

The KJNodes prompt builder separates the global prompt into fields such as:

• high-level scene description;
• background and visible environment;
• style category;
• photographic or illustrative treatment;
• subject aesthetics, clothing, and appearance;
• lighting; and
• medium or shot type.

Keep these fields consistent. A prompt that asks for a candid smartphone image in one field and a formal studio composition in another gives the model conflicting guidance.

Draw object boxes

Use object boxes for a subject or prop whose position matters. Each box stores normalized coordinates, dimensions, a type, and a text description. Describe the object itself rather than adding unrelated scene instructions.

For example:

woman seated at the table in a brown strapless dress

A smaller box can target a hand, phone, bag, steering wheel, or another local detail. Avoid dozens of overlapping boxes: regional instructions can conflict when they claim the same space.

Draw text boxes

For visible words, use a text box and enter the exact intended text. Ideogram is designed for text rendering, but spelling, logos, and typography still require review. Do not assume that a generated brand mark is accurate or licensed for publication.

Inspect the generated JSON

Run or preview the prompt builder before the full sampler when possible. Check that:

• the scene summary matches the intended image;
• boxes stay within the canvas;
• text strings are exact;
• object descriptions do not contradict the global prompt; and
• palettes use valid colors.

Important generation settings

The supplied workflows record these starting values:

The video suggests reducing the sampler from 28 to 20 steps for faster iteration. This is a demonstrated option, not a verified universal optimum. Compare results on your own prompts before standardizing on a lower step count.

Changing the seed produces a new variation while keeping the prompt and layout controls. If a composition is close but the face, hands, or small objects are weak, test a few seeds before rewriting the entire prompt.

About the Extend Intermediate Sigmas node

The supplied graphs include an ExtendIntermediateSigmas node with saved values 2, 1, 0.98, and linear. The video describes this as bypassing safety filters and making the model “100% uncensored.” That claim is not verified.

Official ComfyUI documentation says Ideogram 4 has a built-in safety filter and notes that structured JSON can reduce false positives. Treat the sigma modification as an experimental community workflow setting, not a guarantee that filtering is removed or that every request will generate.

Troubleshooting

Missing Ideogram or prompt-builder nodes

Update ComfyUI, KJNodes, RES4LYF, and rgthree-comfy. Install ComfyUI-GGUF if you imported the NVFP4 file. Restart ComfyUI and reload the JSON.

Model not found

Check the exact filename and folder. The conditional and unconditional files are separate; both are required. Do not place the text encoder in diffusion_models or the VAE in checkpoints.

NVFP4 produces incorrect loader errors

Open the two UNET loader nodes. Select:

• conditional: ideogram4_nvfp4_mixed.safetensors
• unconditional: ideogram4_unconditional_nvfp4_mixed.safetensors

The packaged NVFP4 JSON does not currently make this distinction correctly.

Image blocked by safety filter

Rewrite the request as valid structured JSON and remove ambiguous wording. ComfyUI's official guide says plain-text prompts have a higher false-positive rate. Do not assume that changing sigma settings reliably disables the built-in filter.

Bounding boxes are ignored or misplaced

Simplify the layout, reduce overlaps, and make each regional description concrete. Check that the box matches the selected resolution and that global text does not place the same object elsewhere.

Out-of-memory errors

Start near 1 megapixel, close other GPU workloads, and confirm that you loaded the intended precision. NVFP4 is intended for compatible RTX 50-series hardware; FP8 may be more appropriate on other supported NVIDIA GPUs. Actual memory needs vary, so avoid treating file size as peak VRAM usage.

Limitations and trade-offs

• The license is non-commercial. Open weights are available, but commercial deployment and monetized output use are restricted by the license.
• Layout guidance is not a guarantee. Boxes improve control but do not force exact geometry in every seed.
• Text still needs proofreading. Ideogram emphasizes text rendering, but generated spelling, logos, and type can still be wrong.
• Two large denoisers are required. The conditional and unconditional models increase storage and loading requirements.
• Workflow metadata is inconsistent. The NVFP4 file contains stale FP8 download metadata and an incorrect unconditional-loader selection.
• Performance claims are hardware-specific. The video's sub-two-minute and 24 GB observations are not universal benchmarks.
• Leaderboard positions change. The narration's Elo and ranking claims were not reproduced from a matching dated leaderboard snapshot and should not be used as permanent product claims.
• Safety behavior remains. The official workflow documents a built-in safety filter; “100% uncensored” is not established.

FAQ

Is Ideogram 4 fully open source?

It is more precise to call Ideogram 4 open weight. Its weights and inference code are public, but the weights are governed by the Ideogram 4 Non-Commercial license rather than a permissive license such as Apache 2.0 or MIT.

Does Ideogram 4 require JSON prompts?

No. Official ComfyUI documentation supports natural language and structured JSON. JSON is preferable when you need bounding boxes, palettes, explicit text regions, or detailed layout control.

Which workflow should I use?

Use FP8 for the standard supplied setup on compatible NVIDIA hardware. Use the NVFP4 workflow only on compatible RTX 50-series hardware and correct its unconditional model loader first.

Is the Gemma 4 text encoder required?

Not for the supplied FP8 graph as currently wired. Its note lists a Gemma 4 encoder, but the graph loads Qwen3-VL only. A newer official template may use a different encoder configuration.

Can I use Ideogram 4 for paid AI-influencer work?

Do not assume so. The model license defines non-commercial purposes and specifically restricts output used in or to promote revenue-generating products or services. Review the license and obtain appropriate permission before commercial use.

Conclusion

The supplied Ideogram 4 ComfyUI workflows provide a practical interface for the model's structured prompting and bounding-box controls. The FP8 graph is the cleaner starting point; the NVFP4 variant offers smaller model files for compatible RTX 50-series cards but requires a loader correction and careful review of its stale notes.

Start at a modest resolution, inspect the generated JSON, use a small number of non-overlapping boxes, and treat the model's license and safety behavior as real constraints. Those details matter more than leaderboard slogans when deciding whether the workflow fits a production pipeline.