Fixing OpenAI’s color bias with simple RGB scaling

| May 8, 2025

A quick fix to counteracting the warm bias in GPT-4o image generation.

Table of contents

When AI paints too warm

Imagine you’ve asked GPT‑4o for a crisp and professional image, only to see every edge bathed in buttery yellow. That’s what greeted users when OpenAI unleashed its new multi‑modal image model in March 2025. Headlines and social feeds filled up with extremely detailed characters, lush forests, and even scenes inspired by Studio Ghibli, yet each image shared the same telltale warm cast.

In this post, we’ll first quantify GPT‑4o’s warm bias by asking the model to reconstruct known reference images, then introduce a single‑line RGB‑scaling hack that restores true‑to‑life color across every pixel. Let’s dive in.

Fixing OpenAI’s color bias with simple RGB scaling - Figure-1 — Figure 1: Most images display a noticeable yellow bias

Building a dataset

Because GPT‑4o invents every pixel from scratch, it offers no built‑in “ground truth” for color accuracy. To quantify the yellow shift, we supply it with a set of high-quality refined images and ask it simply to replicate them. The striking result: even when told to produce an exact copy, reconstructions consistently lean warm, confirming that the tint is a model artifact, not a prompt quirk.

Fixing OpenAI’s color bias with simple RGB scaling - Figure-2 — Figure 2: Original vs reconstructed image. The reconstruction process tends to keep a similar image while showing warmer tint.

Every regenerated image in our canonical set exhibits the same yellow cast (Figure 2), demonstrating that the bias persists under controlled reproduction. With this test set in hand, we can now measure channel‑by‑channel shifts and design a corrective remedy.

Measuring color bias

Although reconstructions mirror the originals in overall composition, subtle changes do occur: objects may slightly shift in shape or position, for example. Yet these shape differences are minor compared with the systematic color drift revealed by per‑channel averages. The next figure plots mean‑pixel differences for red, green, and blue (reconstructed minus original), exposing two key patterns.

Fixing OpenAI’s color bias with simple RGB scaling - Figure-3 — Figure 3: Distribution of RGB shifts (reconstructed vs. original)

First, reconstructions exhibit a slight darkening across all channels. More critical is the uneven drop in blue intensity: blue falls significantly more than red or green, yielding the warm, yellow‑tinted cast we observed visually. By quantifying this imbalance, we can identify shifts in our images and boost some channels to better match the color of the original image.

Correcting via RGB scaling

One straightforward idea was to add a constant value to each channel (boost blue by +12, red by +5, and so on) to counteract the average shifts we observed. In practice, this “RGB bias” tended to over-correct images and under‑correct others, making dark images too bright, while others saw almost no change.

Why scaling works better

Instead of adding, we scale each channel by a factor $\alpha_c$, like gently turning up the blue knob on a mixer by a percentage. This preserves an image’s natural contrast: shadows remain deep, highlights stay crisp, and colors regain their intended balance.

The core formula

We apply a simple formula, pixel by pixel:

Fixing OpenAI’s color bias with simple RGB scaling - Formula

Here, each channel c (red, green, blue) is multiplied by its own scaling factor α_c.

Finding the sweet spot

We swept each α_c across a range and plotted reconstruction error vs. boost factor (see the figures below). The clear troughs in these curves show exactly where each channel is best corrected, no guesswork needed!

Fixing OpenAI’s color bias with simple RGB scaling - Figure-4 — Figure 4: Reconstructed error for each channel

Results: before & after correction

Below is a direct comparison: on the left, GPT‑4o’s uncorrected output suffers from a warm, yellow cast; on the right, our single‑step RGB scaling restores true-to-life blues and neutral tones.

Fixing OpenAI’s color bias with simple RGB scaling - Figure-5 — Figure 4. Left: original GPT‑4o render with warm bias. Right: after applying per‑channel scaling. Notice the cooler shadows, balanced highlights, and neutral midtones.

Get started: try it on Freepik

Ready to see the difference for yourself? Head over to Freepik and generate your next set of images with true-to-life color every time. With the Freepik Assistant, you can not only enhance your photos using tools like Color-Fix, but also take your visuals further: iterate on the improved results to explore new styles or compositions or even transform them into short videos with just a few clicks. It’s all part of a seamless, AI-powered workflow designed to help you get more from every image. And if you’d like a quick standalone test, you can also experiment with the Color‑Fix tool directly here.

By Freepik