New on the blog: why AI belongs in the feed →
View Categories

Troubleshoot Image Generation

Target File Upload Field Error #

The destination must be an existing File Upload field on the same form. For after-submission generation, it must not be required. Save form changes and reload the feed editor if the selector is stale.

Unsupported Size Or Quality #

Use a listed size for the selected model. A custom size is supported only for GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, and GPT Image 2, and must satisfy every displayed dimension rule. The Extra High and Maximum quality levels are supported by GPT Image 2.5 and 2.6 Flare and Sunburst. A custom model ID receives neither allowance, however similar its name looks, because the plugin only grants them to the model IDs it has reviewed.

Automatic size and quality are provider choices. If the resulting dimensions or quality must be predictable, choose explicit supported values.

GIF Generation And Complexity #

Animated GIF requires a listed GPT Image 2.6 model and GIF preview access on the OpenAI project. If OpenAI rejects the frame/size combination, choose Automatic size or reduce the frame count. The provider enforces the combined pixel budget; the plugin validates frame count and timing before making the request. Frame Count must be 2–16; Frame Duration must be 20–2,000 milliseconds and is rounded to 10 milliseconds. Frame controls are ignored for still output formats.

Complexity is supported only by the reviewed GPT Image 2.6 models. Leaving it at the provider default uses High. If generation is slow, try Low or Medium complexity separately from Quality.

Provider Rejects The Prompt #

OpenAI can decline a prompt under its content policy. The feed reports this as “The image provider declined this request under its content policy,” and no image is produced.

What The Visitor Sees #

The visitor is only ever asked to change their own answers. When the wording that caused the refusal came from the feed’s prompt, they are told to contact the site administrator instead, because there is nothing in their answers for them to fix. When the cause cannot be pinned down, the subject is named so they can check for it themselves.

You can replace the opening sentence with the Content Refusal Message setting on the feed.

Where To Look #

GravityOps -> AI Feed -> Overview lists recent refusals under “Recent image refusals”:

  • Origin — whether the wording came from the feed prompt, the visitor’s answers, or the generated image.
  • Likely cause — the subject OpenAI objected to, such as children, weapons, or lifelike people.
  • Request ID — quote this if you raise a refusal with OpenAI.

A refusal marked Feed prompt affects everyone using the form and will keep happening until the prompt is edited. Nobody will report it, because visitors are never shown that cause.

The image feed list also carries a Refusals (30d) column, so a feed that is quietly failing is visible beside the feed itself.

Retrying #

OpenAI checks the prompt before generating and the finished image afterwards. If the prompt was refused, retrying it unchanged always fails, so something has to change. If only the finished image was refused, retrying can succeed on its own, and the message says so.

Where the prompt needs to change, work through it in this order:

  1. Simplify the prompt. Photorealistic depictions of identifiable people are the most frequent cause. Describing a setting, an object, or an activity is usually accepted where describing a person in difficulty is not.
  2. Check the merged values. Every merge tag is sent to OpenAI, so a visitor’s answer can introduce content the prompt’s author never saw. Confirm the merge tags resolve to the values you expect, and that no tag is being sent unreplaced as literal text.
  3. Consider the Content Filter setting. Switching it to Less restrictive asks OpenAI to relax filtering to the extent it permits. This reduces refusals; it does not eliminate them, and OpenAI’s policies still apply.

Because refusals cannot be ruled out, any form that requires a generated image needs a route forward when one does not arrive. The target of an image feed is always a File Upload field, so tell visitors they may upload a file themselves; a required target field with no such instruction leaves them unable to continue.

Preview Appears But Entry Has No File #

Frontend generation creates a temporary verified session. The form must subsequently submit with the matching form/feed/session context, and the target field/feed configuration must remain consistent. Confirm the visitor did not leave the page, submit a different form, or use stale cached markup from before the feed changed.

If a single-file target also contains a manual upload, the plugin blocks generation before the provider request. If a file is selected after generation, submission stops with an instruction to remove the existing file. This replaces Gravity Forms’ less actionable “Number of files (2) exceeds limit (1)” error. Ordinary validation failures preserve the verified session and restore its image preview; the temporary session still expires after one hour.

Image Is Slow Or Expensive #

Lower quality, use a smaller listed size, simplify the workflow, and limit public requests. Review current OpenAI project usage and pricing. Repeated clicks and administrator reruns create new image requests.

The Generated SVG Was Not Saved #

The provider returned an SVG that contained something other than static drawing content, such as a script, a link, an embedded image or animation. The plugin rejects such a file instead of saving a changed version of it. The error names what was found. Generate the image again; if it keeps happening, simplify the prompt or choose PNG.

If OpenAI rejects an SVG request because of the svg parameter, check the feed’s SVG Author Model. It must be filled in when SVG model selection is enabled on your OpenAI project and left blank when it is not.

If the form shows “this file type is not permitted” for a generated SVG, the target File Upload field probably limits file extensions. Add svg to its allowed list.

Image Editing Fails #

The entry note names the cause. The common ones:

  • No source image was uploaded for this entry. The Image To Edit field was empty.
  • The markup overlay must be the same size as the source image. Export the overlay at the photo’s exact pixel size. A phone photo may be stored rotated; use the size it displays at.
  • The markup overlay must be a PNG with a transparent background. The marks were drawn on a copy of the photo, or the file was saved without transparency.
  • The source image must be a PNG, JPEG, WEBP image. Other formats, such as HEIC, are not accepted.

File Format Problems Downstream #

Choose PNG, JPEG, WebP, or GIF based on the systems consuming the upload. If another add-on rejects WebP or expects JPEG, change the feed’s Output Format and generate a new image; changing the setting does not convert files already attached to entries.

GIF previews and attachments retain the original animation bytes. If an external system displays only a still frame, check whether that system resizes or converts GIF uploads.

Stuck? Ask in GravityOps Help #

Search past answers or post your question, tagged "ai-feed".