OpenAI Compounds

Introduction

OpenAI compounds in Aximmetry let you generate and edit images, chat with OpenAI models, and generate video clips directly in Flow Editor graphs.

This page explains how to set up OpenAI and how to use the following four OpenAI compounds:

• OpenAI_Text-ChatGPT: Runs conversational text workflows. Depending on the selected model, it can also accept image input and return generated images.
• OpenAI_Image-Generate: Generates a new image from a text prompt.
• OpenAI_Image-Edit: Edits or creates variations of an existing image from a text prompt.
• OpenAI_Video-Generate: Generates a video clip from a text prompt.

Set Up OpenAI and API Keys

Before you use any OpenAI compound, you need a working OpenAI account and an API key. You can use the same key with all four compounds.

Create an OpenAI Account and API Key

1. Open the OpenAI Platform and sign in or create an account.
2. Open the API Keys page.
3. Click on Create new secret API key.
   
   NOTE: Consider creating separate keys per project or team member to simplify access control.
   NOTE: If you create a restricted key, allow the endpoints required by the compounds you use. In practice, this usually means Responses (within Model capabilities group) for Text-ChatGPT, Images (within Model capabilities group) for image generation or editing, and Videos for video generation.
   
   
4. Copy the key and store it in a secure place (for example, a password manager).

IMPORTANT: The API key is sensitive. Do not share it, and do not include it in scenes or compounds that you distribute.

Connect the API key in Aximmetry

The OpenAI compounds are located under [Common]:Compounds\AI\.

1. Open your project in Aximmetry and add the OpenAI compound you want to use to your Flow Editor graph.
2. Select the compound and locate the OpenAI API Key input pin.
3. Paste the key into the pin value:
   
   NOTE: If you plan to share the project, do not store a real API key in the scene. For example, use a Text File module to read the key from a secure file and connect it to the OpenAI API Key pin.
   
   
4. Start the compound. If the key is valid, the Model pin updates with the available models and is no longer set to Auto.

Model Access, Usage, and Rate Limits

Your OpenAI account must have access to the models you select on the pins. Requests can fail if a model is unavailable for your account, if you reach your usage quota, or if you hit rate limits.

• Use each compound’s Status output to see error messages returned from OpenAI.
• Avoid triggering many requests in a short time in production scenes to reduce the chance of rate-limit errors.

Shared Concepts for OpenAI Compounds

The four OpenAI compounds share several core concepts. Understanding them helps you use the compounds more effectively and explains how they interact with the OpenAI API.

Prompts and Instructions

Each compound has a text input pin that describes what you want the model to produce.

• In the Image and Video compounds, this text is the prompt (instruction) for the visual output. The image or video generation model processes it directly. These models have limited reasoning capability and rely mainly on associations.
• In the ChatGPT text compound, this text is the message you send to the model. Chat models can follow more complex instructions and perform reasoning.
  In most cases, use the base model ID (for example, gpt-5.2) rather than a specialized variant (for example, gpt-5.2-codex or gpt-5-nano)

NOTE: When you use the ChatGPT compound for image generation, it rewrites your message into an image-ready prompt in the background and sends that prompt to the image generation model.

If your prompt is longer than a few words, connect a Text Block module instead of typing the text directly into the pin.

Models and Model Selection

Each compound has a pin that selects the OpenAI model family or a specific model. The compounds fetch model lists from OpenAI and expose them on dropdown pins.

• Model Family and Model for OpenAI_Text-ChatGPT.
• Image Model for OpenAI_Image-Generate and OpenAI_Image-Edit.
  
• Video Model for OpenAI_Video-Generate.

Some settings, such as quality, resolution, duration, or reasoning effort, are only valid for certain models.
Models with Pro in their name cost substantially more than other models.

IMPORTANT: Start with lower-cost models and lower-quality settings while you refine your prompt. This is especially important for video generation, where cost and render time can increase quickly.

Resolution, Duration, and Quality

Image compounds expose Resolution and Quality. The video compound exposes Resolution and Duration. These settings strongly affect cost, latency, and usability.

• Use lower resolutions and shorter durations while you experiment.
• Match the resolution to your scene framing and output aspect ratio.

Output Folders, Undo, and History

Images and videos are saved to the folder pins provided by each compound, such as Output Folder or Image Folder. The saved file paths are exposed on outputs such as Saved Image Path(s) and Saved Video Path. Use these paths to archive results or pass them to other file-based modules.

All four compounds offer an Allow Undo input. When it is enabled, Aximmetry’s Undo and Redo operations also restore the compound’s generated outputs, so the Flow Editor history and media stay in sync.

NOTE: When you use Input Image, the sent input images are saved to Documents:Recordings.

Status and Completion

Use the compound's trigger pin to send the request to the OpenAI API. Depending on the compound, this pin is Send, Generate Image, Apply Edit, or Generate Video.

While the compound is waiting for results, it appears brighter and more saturated than normal:

Every OpenAI compound exposes:

• A Status text output with success and error information for the last request.
• A Finished trigger output that fires when the request completes, and outputs are updated.

Use these outputs to update user interfaces, trigger follow-up processing, or show clear error messages to operators.

OpenAI_Text-ChatGPT

OpenAI_Text-ChatGPT sends prompts to OpenAI chat models and returns text responses. Depending on the selected model and enabled options, it can also analyze input images, generate images, and use web search.

Use it when you need conversational logic, text and data processing, or text-driven control flows in Aximmetry.

For example, you can:

• Build on-air or in-scene assistants that answer questions.
• Generate captions, summaries, or even diagrams based on operator prompts or external data sources such as RSS feeds.
• Interpret text commands and map them to scene controls.
• Combine text responses with generated images in a single workflow.

Input Pins

• OpenAI API Key: Provides the secret API key for chat requests. Keep this value private.
• Prompt: Contains the message you send to the chat model. If Auto New Chat is disabled, this text is appended to previous exchanges. Connect a Text Block module for longer prompts.
• Model Family: Filters the list of models shown by the compound so you can focus on the latest models.
• Model: Selects the specific chat model to use. In most cases, use the base model ID (for example, gpt-5.2) rather than a specialized variant (for example, gpt-5.2-codex or gpt-5-nano)
• Reasoning Effort: Controls the reasoning effort for supported models. Higher effort can increase latency and cost.
• Auto New Chat: Starts a new conversation automatically when you trigger Send.
• Set New Chat: Clears previous prompts and responses when triggered.
• Aximmetry Context: Contains Aximmetry-specific context (system prompt) that helps the model respond in an Aximmetry-aware way. The text is stored at [Common]:Compounds\AI\Elements\AximmetryContext_Text.txt.
• Allow Web Search: Let the model use a web-search tool to search on the internet for information.
• Image Generation: Enables image-generation capabilities in the chat workflow when supported.
• Input Image Mode: Enables the Input Image pin.
• Input Image: Provides an image frame for the model to analyze. The frame is sampled when you trigger Send.
• Allow Undo: Includes chat outputs and any generated images in the Undo and Redo history of the Flow Editor.
• Image Out Pin: Selects which image is output through the Image Out pin when multiple images are available. If All (Stitched) is selected, all received images are stitched together and output side by side through the Image Out pin.
• Send: Sends a single request to the selected chat model.

Output Pins

• Response Text: Contains the primary text answer from the model. This output can be empty when the last result only returned images.
• Image Out(s): Displays the decoded image or images returned in the response. For multi-image results, it behaves like other image compounds. Use the compound's Image Out Pin input pin to select which image to display.
• Saved Image Path(s): Lists file paths of any images saved during the response, one path per line.
• Finished: Fires when the response is complete, and outputs are updated.
• Status: Reports HTTP and tool-level errors or success information.

OpenAI_Image-Generate

OpenAI_Image-Generate generates one or more images from a text prompt using OpenAI image models.

Use it when you need still images that you can treat as textures, backgrounds, or design elements in your scene.

For example, you can:

• Create graphic elements or backgrounds for virtual sets.
• Explore visual ideas and styles for a production.
• Generate placeholder images while you design a workflow.
  
• Generate multiple image variations of a concept for your audience during live production.

Input Pins

• OpenAI API Key: Provides the secret API key for image generation requests. Keep this value private.
• Prompt: Describes the image or images that the model should generate. Connect a Text Block module for longer prompts.
• Image Model: Selects the image model from the list returned by OpenAI.
• Quality: Controls the level of visual detail in the generated images. Some quality levels are only valid for certain models.
• Resolution: Defines the width and height of the generated images. Use portrait or landscape to match your scene layout.
• Number of Images: Sets how many images to request in a single call. Higher values increase cost and processing time.
• Background: Chooses the background style for the generated images. Transparent backgrounds require a model that supports alpha output.
• Moderation: Controls the strictness of moderation for image requests. At the default auto setting, the API refuses requests that contain disallowed content, such as graphic sexual content, graphic violence, or hateful content. When set to low, filtering is less restrictive. 
• Output Folder: Specifies where generated image files are stored on disk. Ensure that the folder is writable.
• Image Out Pin: Selects which result appears on the image output when multiple images are returned. You can select a single image index or the All (Stitched) option, which outputs a tiled contact sheet of up to nine returned images.
• Allow Undo: Integrates generated outputs with the Undo and Redo history of the Flow Editor.
• Generate Image: Triggers a new image generation request using the current settings.

Output Pins

• Image Out(s): Provides the decoded image result of the request. For multiple images, this output can show one selected image or a tiled contact sheet. Use the compound's Image Out Pin input pin to select which image to display.
• Saved Image Path(s): Lists the file paths of saved image outputs, one path per line.
• Finished: Fires when the request is complete and the image output is updated.
• Status: Describes the result of the last request, including error messages from OpenAI.

OpenAI_Image-Edit

OpenAI_Image-Edit edits or varies an existing image based on an instruction and an optional mask.

Use it when you want to refine generated images, adjust photographs, or extend content beyond the original frame.

For example, you can:

• Retouch images or remove unwanted details.
• Replace objects or regions according to a mask.
• Extend a background to match a new framing.
• Iteratively refine an image by reusing the previous result.

Input Pins

• OpenAI API Key: Provides the secret API key for edit requests.
• Instruction: Describes how the model should change or vary the Input Image.
• Input Image: Supplies the source image that will be edited. The frame is captured when you trigger the Apply Edit pin.
• Mask Image: Limits edits to selected areas of the image. Bright regions are treated as editable, and dark regions are preserved, depending on the model.
• Use Image Out: Reuses the previous result as the next input image so you can build iterative image-to-image workflows.
• Image Model: Selects the image model used for editing.
• Quality: Controls the level of detail in the edited result.
• Resolution: Sets the output resolution of the edited image.
• Number of Images: Defines how many edited variants to request in one call.
• Background: Chooses the background style for edited results. Transparent backgrounds require model support.
• Moderation: Controls moderation strictness for edit requests. At the default auto setting, the API refuses requests that contain disallowed content, such as graphic sexual content, graphic violence, or hateful content. When set to low, filtering is less restrictive. 
• Output Folder: Specifies where edited image files are saved. Ensure that the folder is writable.
• Image Out Pin: Selects which result appears on the image output when multiple images are returned. You can select a single image index or the All (Stitched) option, which outputs a tiled contact sheet of up to nine returned images.
• Allow Undo: Integrates edited outputs with the Undo and Redo history of the Flow Editor.
• Apply Edit: Triggers a new edit request using the current inputs and settings.

Output Pins

• Image Out(s): Provides the edited image result. For multiple images, this output can show one variant or a tiled contact sheet. Use the compound's Image Out Pin input pin to select which image to display.
• Saved Image Path(s): Lists the file paths of saved edited images, one path per line.
• Finished: Fires when the edit request completes, and outputs are updated.
• Status: Reports success or error details from the last edit request.

OpenAI_Video-Generate

OpenAI_Video-Generate generates video clips from a text prompt using OpenAI video models and plays them back through an internal video player.

Use it when you need to generate moving footage inside Aximmetry. Video generation can take noticeably longer than text or image generation.

For example, you can:

• Create storyboard or animatic clips for planning shots.
• Produce background or B-roll footage for virtual production scenes.
• Explore motion concepts before committing to full production.
• Create animated textures or infographics.

Input Pins

• OpenAI API Key: Provides the secret API key for video generation requests.
• Prompt: Describes the video you want the model to create. Include camera motion, subject, and style where relevant.
• Video Model: Selects the video model from the list returned by OpenAI.
• Duration: Sets the requested video length in seconds.
• Resolution: Defines the output resolution of the generated clip.
• Output Folder: Specifies where the generated video file is saved.
• Use Audio Device: Routes audio from the internal video player to the selected Audio Device.
• Audio Device: Selects the playback audio device used when the Use Audio Device audio routing is enabled.
• Allow Undo: Integrates generated videos with the Undo and Redo history of the Flow Editor.
• Generate Video: Starts a new video generation request.
• Remix: Requests a new video take based on the previous result and the current prompt. Settings such as resolution or model changes are ignored during remix.

Output Pins

• Video Out: Provides video playback output from the internal player. Connect this to your program or preview chain.
• Audio Out: Provides audio playback output from the internal player. Route this into your audio mixing modules.
• Saved Video Path: Contains the file path of the saved video clip.
• Finished: Fires when the video request completes, and outputs are updated.
• Status: Reports success or error information from the last request.

Examples

Weather Forecast Data Example

This example shows how to generate a simple weather forecast graphic dynamically.

In the Text Block module, enter a prompt similar to the following:

The Text Concat module appends the weather data returned by the HTTP Request module below the prompt from the Text Block module.

This example uses the MET Weather API to retrieve weather data for London. Set the URL pin of the HTTP Request module to: https://api.met.no/weatherapi/locationforecast/2.0/compact?lat=51.5074&lon=-0.1278

Because the prompt requests three images, the model returns three variants. Set the Image Out Pin input to the variant you want to display from Image Out(s). If you need a more consistent result, use a more specific prompt or request additional variants.

The model interprets the weather data returned by the MET Weather API to generate the weather forecast graphic based on the prompt

Finally, connect the selected image output to a virtual screen and press Send in the OpenAI_Text-ChatGPT compound.
In this example, Image Out(s) is connected to the VIDEO input of the LCD Screen compound in the News Room example scene: [Studio]:News Room\News Room - VirtualCam_3-Cam.xcomp

NOTE: The general-purpose gpt-5.2 text model was used in the above example.

Lower-Third Image Example

This example creates a lower-third graphic for an Easter broadcast.

Set Prompt to a text such as: “Create an Easter bunny as a lower-third graphic for an Easter-themed news broadcast.” This is still a general prompt. In most cases, you will get better results with a more detailed prompt.

In this example, use an image model such as gpt-image-1.5 with Quality set to High and Resolution set to 1536x1024, which fits a landscape production. Use a Transparent Background so the graphic can be composited over the final video as a lower third.

Generate three images with the Number of Images set to 3.
Review them in the preview, then select the best one with the Image Out Pin in the OpenAI_Image-Generate compound. Then use the Placer module’s Offset and Scale pins to position the graphic on the screen.

You can turn the Easter lower third on and off in the final Out by enabling or disabling the Blender 2 module through its Active pin, which is connected to the Copy Logic module in the first screenshot.

Improving Results

Better Prompting

In most cases, a more detailed prompt produces more consistent results and better matches your intended style or idea.

You can also use the OpenAI_Text-ChatGPT compound to expand a short prompt into a more detailed one. In this example, the prompt is to "Improve this prompt, make it more detailed:...".

After you receive the improved prompt in Response Text pin, connect it to a Text Block module and use Capture Input on the Text pin to save the result into the Text Block.

After that, disconnect the Text pin from OpenAI_Text-ChatGPT, click on the Edit Text to open the stored text in the Text Block, and edit it further as needed.

Modifying an Image

If the returned image is close to the result you want, but still needs a few changes, use the OpenAI_Image-Edit compound.

In this example, the only change is replacing the text “Easter News” with “Aximmetry News,” so that instruction is set on the compound:

The image model correctly detected the placement of the text and edited it:

NOTE: You can turn on Use Image Out in OpenAI_Image-Edit to continue editing the returned image with new additional instructions.
NOTE: In some cases, the models will only touch a specific region. But in most cases, they will slightly edit the whole image, these edits are usually only noticeable on closer inspection.

Robot Talent Example

This example creates a virtual talent: a robot dog.

The following prompt was used: “Create a realistic robot dog, centered and fully visible in frame for the entire shot, against a seamless chroma key green screen background, with even studio lighting, a static camera, and no cropping.”

This wording is deliberate. Because the video model does not output transparent pixels, the prompt instead asks for a seamless green-screen background that can be keyed out with Aximmetry’s keying tools. The rest of the prompt helps keep the full character visible and the camera static.

While refining a prompt for this use case, it is practical to use sora-2 instead of sora-2-pro to reduce cost. It also helps to keep the duration and resolution as low as possible during testing.

The result is shown below.

In the generated result, the robot dog's back legs start to hover. Because the prompt does not specify what the dog should do, the model generates motion on its own. If needed, continue iterating with Remix and ask the model to keep its legs on the ground. If that still does not produce the result you want, generate a new video.

Troubleshooting

Rate Limits, Quota, and Model Availability

Symptoms

• The Status output pin mentions rate-limit, quota, or model availability errors.
• Requests sometimes succeed and sometimes fail during intense usage.

Cause

Your OpenAI account or project has reached a rate limit or usage limit, or the selected model is unavailable for your usage tier or organization verification status.

Fix

• Send fewer requests in a given time period.
• Reduce the number of images, resolution, or duration.
• Try a different model if the current one reports availability issues.
• Design flows so they only trigger new requests when outputs are needed.
• If necessary, increase your usage limit or request a higher limit in your OpenAI settings.

Slow Responses and Timeouts

Symptoms

• Requests take a long time to complete.
• The Status output pin reports timeout or network errors.

Cause

Large inputs, long durations, high request volume, or increased reasoning can increase latency. Some higher-compute models can take several minutes to finish, and upstream timeout, proxy, or network issues can cause requests to fail before they reach OpenAI.

Fix

• Use lower resolutions, fewer images, or shorter durations.
• Disable features such as web search or high reasoning effort when you do not need them.
• Reduce reasoning effort if the selected model supports it.
• Check your internet connection and any proxy or firewall configuration.

Nothing Returned

Symptoms

• The Status output pin and the Response output pin return nothing, even though the request appears to have finished.

Cause

In rare cases, the model may return no visible output. This can also happen if the prompt leads to an empty response.

In the text compound, Response Text can also be empty when the last result returned only images.

Fix

• Try the request again.
• If the issue persists, check the OpenAI dashboard Usage views. The OpenAI_Text-ChatGPT request should also show up in the Responses Logs. The video results should show up for 60 minutes in the Videos view's history.