Ideogram 4 for Forge Neo

This extension adds a separate Ideogram 4 tab to Forge Neo and runs Ideogram's official inference package in an isolated Python environment. Forge's own Torch and dependencies are not modified.

Install from Git

1. Accept the gated model license: https://huggingface.co/ideogram-ai/ideogram-4-nf4

2. In Forge Neo, open Extensions > Install from URL.

3. Paste this repository's Git URL and install it.

4. Restart Forge Neo. The extension creates its isolated environment and installs its dependencies during startup. The first setup downloads a large CUDA-enabled PyTorch package and can take several minutes.

5. Either paste a Hugging Face read token into the tab or authenticate the extension environment:

   env\Scripts\hf.exe auth login

6. Open the Ideogram 4 tab.

For a manual installation, clone this repository into Forge Neo's extensions directory and restart Forge Neo:

cd path\to\sd-webui-forge-neo\extensions
git clone https://github.com/Whatwhatio/forge-neo-ideogram4.git ideogram4_forge

If automatic setup fails, close Forge Neo and run setup.bat once.

The first generation downloads the weights to the normal Hugging Face cache and loads the model automatically. Download, model-loading, sampling-step, decoding, and save progress appears in both the Forge Neo console and the tab.

The blue down-arrow restores the settings from the previous generation attempt. The trash button clears the prompt and restores builder defaults. Settings are saved before generation begins, so they survive a failed generation or worker crash. Hugging Face and Ideogram API tokens are never written to this file.

Download full-page screenshot captures the complete Forge page without changing browser zoom. The PNG Info tab reads metadata from Ideogram PNGs and can restore their prompt, settings, and boxes into the visual builder.

Visual JSON Builder

You can write an Ideogram JSON caption manually in the Prompt tab, but this is not recommended for most users. The required structure, key ordering, bounding boxes, and nested style fields are much easier to manage with the Visual JSON Builder.

Open Visual JSON Builder and check Use visual builder for generation. When enabled, the extension constructs the complete Ideogram caption from the form and layout boxes.

The example above describes a painterly anime-style scene of a woman playing an acoustic guitar in a Mediterranean alley:

• High-level description: Summarize the complete image in one or two sentences. Include the main subject, action, location, composition, and overall visual style. In the example, this describes an adult brunette woman playing an acoustic guitar in a sunlit Mediterranean alley.
• Background / environment: Describe only the setting that should fill the whole image behind the positioned subjects. Include architecture, landscape, weather, lighting, and environmental details. In the example, this contains teal-painted doors, weathered walls, stone buildings, balconies, foliage, steps, and sunlight. Do not repeat the main subject here.
• Style type: Choose Photograph for realistic photographic output or Artwork for anime, illustration, painting, graphic design, and other non-photographic styles.
• Aesthetics: Describe the overall visual character and mood, such as expressive anime illustration, bold painterly brushwork, saturated teal and golden-orange contrast.
• Lighting: Describe the direction, color, quality, and intensity of the light. The example uses intense late-afternoon sunlight, warm rim light, cool turquoise shadows, and reflected light.
• Medium: Enter the physical or digital medium, for example photograph, digital painting, watercolor, or 3D render.
• Camera details or art style: For photographs, describe lens, viewpoint, depth of field, and photographic treatment. For artwork, describe rendering technique, linework, brushwork, shading, and finish.
• Overall palette: Add optional hexadecimal colors separated by spaces, such as #073F4B #087F88 #F49A18 #FFD66B. These guide the image-wide color scheme.

Layout boxes

Use boxes for foreground subjects, props, furniture, signs, and literal text. Drag a box to move it and drag its lower-right handle to resize it. Boxes may overlap. The output Width and Height controls determine the canvas aspect ratio.

• Add object: Adds a region for a person, animal, prop, or other visual object. The example uses box 01 for the woman and box 02 for the guitar. Separating the guitar into its own box gives Ideogram more explicit instructions about its shape and placement.
• Add text: Adds a literal typography region. Enter the exact words in Literal text, then explain the sign, label, material, font, and lighting in Description.
• Delete selected: Removes the currently selected box.
• Show last image: Displays the previous generated image behind the layout boxes for alignment. It is only a preview and is not included in saved or downloaded images.
• Selected box: Selects a box directly when several overlap or when it is difficult to click on the canvas.
• Box color: Changes the editor and output-overlay color for that box. It is display metadata and does not control the generated object's actual color.
• Selected type: Switches the selected region between Object and Text.
• Description: Describe only the contents of the selected region. Include appearance, pose, action, materials, anatomy, and spatial relationships. In the example, box 01 describes the seated woman, while box 02 concentrates on the guitar's strings, neck, bridge, tuning pegs, and hand placement.
• Element palette: Adds optional hexadecimal colors for the selected element only.
• Bounding box [y1, x1, y2, x2]: Fine-tunes the box numerically on an Ideogram coordinate grid from 0 to 1000. The top-left corner is 0,0; the bottom-right corner is 1000,1000.

Generated element layout JSON updates from the boxes automatically. You can also paste compatible layout JSON into this field and click Import JSON into Visual Builder. Importing is useful for prompts shared by other users or elements_data copied from KJNodes' Ideogram4PromptBuilderKJ.

The layout field uses the same normalized x, y, w, h region format as KJNodes' Ideogram4PromptBuilderKJ.

Generation settings

• Hugging Face model repository: Leave this as ideogram-ai/ideogram-4-nf4 unless you intentionally use a compatible repository.
• Hugging Face token: Enter a read token from the Hugging Face account that accepted the model license. Tokens are kept in memory and are not written to the saved prompt settings.
• Sampler preset: Selects Ideogram's official sampling configuration. V4_DEFAULT_20 is the normal starting choice.
• Seed: Use -1 for a new random result each generation, or reuse a specific seed to reproduce and refine a composition.
• Width / Height: Set the output resolution and Visual Builder canvas shape. The example uses 768x1152.
• Ideogram API key for Magic Prompt: Optional and mainly useful for plain text prompts. Leave it blank when using structured Visual Builder captions.
• Allow caption verifier warnings: Allows generation to continue when the official Ideogram verifier reports non-fatal caption-format warnings.
• Generate: Saves the current prompt settings, downloads and loads the model when necessary, then generates the image.
• Blue down-arrow: Restores the previous prompt and generation settings.
• Trash button: Clears the current prompt and returns the builder to its default state.

The first time you click Generate, allow several minutes while the extension downloads the required gated model and its isolated CUDA-enabled PyTorch environment, then loads the model into memory. Later generations reuse the downloaded files and loaded pipeline.

Video example

The linked demonstration uses ComfyUI, but its Ideogram prompt-builder workflow is very similar to this extension's Visual JSON Builder:

Ideogram 4 prompt-builder demonstration in ComfyUI

Safety-filter output

A gray result saying Image blocked by safety filter is not a precise error message. It can mean that Ideogram's safety filter rejected the requested content, or that an incomplete/non-JSON caption caused a false positive. Ideogram's official documentation notes that false positives are more common with captions outside its structured JSON format.

The Generated element layout JSON field contains only the layout elements array. It is valid input for Import JSON into Visual Builder, but it is not a complete caption for the Prompt tab. A complete caption is a JSON object containing compositional_deconstruction, including both background and elements; a high-level description and complete style description are also strongly recommended. Keep Use visual builder for generation checked so the extension wraps the layout boxes in the complete caption structure automatically.

Rollback

Close Forge Neo and delete this entire ideogram4_forge directory. No Forge core files or Forge Python packages are changed.

Licensing

The extension integration code should be distributed under the license chosen by its author. The bundled official_source directory is Ideogram's official inference code and is distributed under Apache License 2.0; its original license is retained in official_source/LICENSE.md.

The Ideogram 4 model weights are not included in this repository. Users download them separately from Hugging Face and must accept Ideogram's Non-Commercial Model Agreement: official_source/model_licenses/LICENSE-IDEOGRAM-4-NON-COMMERCIAL.

This is an unofficial community extension and is not endorsed by Ideogram.