Support custom image dimensions for GPT-Image-2.5 (e.g. 3840x2160)

[kazukata]

Support custom image dimensions for GPT-Image-2.5 (e.g. 3840x2160)

Is your feature request related to a problem?

Yes.

I'm running LibreChat v0.8.8-rc2 and testing GPT-Image-2.5.

I requested a 4K image (3840x2160), but the generated image was only 1365x768. During troubleshooting, I found that the image generation schema appears to restrict the size parameter to the following values:

• auto
• 1024x1024
• 1536x1024
• 1024x1536

GPT-Image-2.5 supports custom dimensions and high-resolution outputs, including 4K-sized images. However, it appears that LibreChat may not currently allow these values to be passed to the API.

Current schema:

size: {
  type: 'string',
  enum: ['auto', '1024x1024', '1536x1024', '1024x1536'],
  description:
    'The size of the generated image. One of 1024x1024, 1536x1024 (landscape), 1024x1536 (portrait), or auto (default).',
},

Describe the solution you'd like

Allow custom dimensions supported by GPT-Image-2.5 to be passed through LibreChat.

Examples:

• 2048x2048
• 2048x1152
• 3840x2160
• 2160x3840

Ideally, LibreChat could support any valid WIDTHxHEIGHT value accepted by the selected image model.

For example:

size: {
  type: 'string',
  description: 'Image dimensions (WIDTHxHEIGHT) or auto',
},

The value could then be validated on the server according to the limits of the selected model.

Describe alternatives you've considered

An alternative would be to extend the existing enum list with commonly used high-resolution sizes:

size: {
  type: 'string',
  enum: [
    'auto',
    '1024x1024',
    '1536x1024',
    '1024x1536',
    '2048x2048',
    '2048x1152',
    '3840x2160',
    '2160x3840',
  ],
},

This would provide access to common high-resolution output sizes while retaining the existing fixed-list approach.

Additional context

Environment:

• LibreChat version: v0.8.8-rc2
• Model: GPT-Image-2.5
• Requested size: 3840x2160
• Actual downloaded image size: 1365x768
• Backend: OpenAI-compatible endpoint

Questions:

1. Is the current JSON schema intentionally limiting the available image generation sizes?
2. If a user requests 3840x2160 or "4K" in the prompt, does LibreChat fall back to size: "auto" because the requested value is not included in the schema?
3. Is 1365x768 the original image size returned by the API, or is the image resized during display, storage, or download?
4. Is support for custom image dimensions planned for GPT-Image-2.5?

Thank you for considering this enhancement.

[danny-avila]

Thanks for the detailed report. Two things are going on, and the second is why you got 1365x768:

1. The schema limits the sizes. The size enum was written for gpt-image-1, so the model can't send 3840x2160. It falls back to auto or one of the listed sizes. 📐 fix: Accept Custom WIDTHxHEIGHT Sizes in OpenAI Image Tools #15904 replaces the enum with a WIDTHxHEIGHT (or auto) pattern for both image generation and editing. The image API then decides which dimensions your model accepts.

2. LibreChat resizes generated images when it saves them. By default, generated images are downscaled to a 768px short side, and 16:9 at that height is exactly 1365x768. So that's our resize on save, not what the API returned.

Workaround for the resize (works today): keep the original resolution by adding this to librechat.yaml:

fileConfig:
  imageGeneration:
    percentage: 100

Or cap the output instead, e.g. px: 3840 limits both width and height to 3840px.

With that setting plus #15904, a 3840x2160 request should be sent as-is and saved at full size.

[oliver-31415]

is there a way to switch between the two new image models that OpenAI released in the gui?