Excessive image compression/downscaling during upload degrades Vision Model performance

[marlonka]

What happened?

Description:
There is a significant issue with how LibreChat handles image uploads. Currently, the application appears to apply heavy compression or downscaling to images before they are sent to the model provider (Vertex AI, OpenAI, etc.). While this might save bandwidth, it effectively "blinds" high-end vision models for enterprise use cases that require fine-detail analysis.

As shown in the attached comparisons, an original high-resolution image (2268x4032, 2.5 MB) is significantly degraded upon upload to img width="768" height="1365" in this case. This makes small text—such as ingredients or price tags on product packaging—completely illegible to the model, leading to hallucinations or "I cannot read this" responses.

Evidence:

• Screenshot 1: Shows a side-by-side comparison. The right side is the original local file; the left side is the image as rendered/stored within LibreChat. The loss of sharpness is immediate.
• Screenshot 2: Highlights a specific vision failure. The original image (right) clearly shows the description for "DRYCK LINGON." The model (Gemini/GPT) struggles to extract precise details because the version it receives is blurred.
• Screenshot 3: Shows the image being served from localhost:3080/images/.... Inspection of the element and the underlying file reveals that the resolution/quality has been stripped down compared to the source file.

Impact:
This is a blocker for enterprise use cases involving:

1. Image Analysis: Small print on forms or labels becomes unreadable.
2. Technical Support: Reading serial numbers or small components on hardware photos.
3. Retail/Inventory: Extracting product data from shelf photos (as seen in the examples).

Expected Behavior:
The system should either:

1. Maintain the original resolution and quality when "Upload to Provider" is selected.
2. Provide a configurable setting in the .env or UI to toggle/adjust the level of image compression.
3. Allow a "Lossless" pass-through for enterprise-grade vision tasks.

Actual Behavior:
Images are aggressively compressed/downscaled, rendering them unsuitable for high-precision vision tasks.

Version Information

LibreChat v0.8.2-rc1(opens in new tab)

Steps to Reproduce

Steps to Reproduce:

1. Upload a high-resolution image (e.g., 4K or >2MB) containing small, sharp text.
2. Use the "Upload to Provider" option with a Vision-capable model (e.g., GPT-5.1 or Gemini 2.5 Flash).
3. Ask the model to transcribe the smallest text in the image.
4. Observe the model failing or providing low-confidence guesses due to the blurred input.

What browsers are you seeing the problem on?

Chrome

Relevant log output

-

Screenshots

uploading via upload to provider to librechat:

see difference:

original image:

after uploading they are resized to smaller resolution and much degraded quality:

Code of Conduct

• I agree to follow this project's Code of Conduct

[Odrec]

Have you tried changing this in the parameter section?

[danny-avila]

This should help if set to high