Too Vivid to Be Real? Benchmarking and Calibrating Generative Color Fidelity

Paper: 2603.10990 Authors: Zhengyao Fang, Zexi Jia, Yijia Zhong, Pengcheng Luo, Jinchao Zhang, Guangming Lu, Jun Yu, Wenjie Pei Categories: cs.CV

The Gap

Text-to-image models like DALL-E and Midjourney can generate stunning visuals, but there’s a subtle problem: when you ask for “realistic” images, they often produce photos that look too perfect, too saturated, too contrasty. Think Instagram filter vs raw camera output. Current evaluation methods (human ratings, preference metrics) actually reward this artificial vividness because humans prefer punchy colors in isolation. But when you need truly photorealistic output—for film, product visualization, or scientific applications—this bias becomes a real limitation.

Problem: Realistic prompts → Overly vivid outputs
    |
    v
Assumption: Current metrics favor visual appeal over authenticity
    |
    v
Method: CFD dataset + CFM metric + CFR refinement
    |
    v
Evidence: 1.3M images with color realism rankings
    |
    v
Conclusion: Framework improves photographic authenticity

The Increment

One sentence: Before this paper, “realistic” image generation was evaluated by visual appeal; after, we can measure and optimize for actual photographic color fidelity.

Core Mechanism

The method has three interlocking components. First, the Color Fidelity Dataset (CFD) contains 1.3M images with ordered levels of color realism—think of it as a spectrum from “obviously synthetic” to “could fool a photographer.” Second, the Color Fidelity Metric (CFM) uses a multimodal encoder trained on this dataset to score how photographically authentic an image’s colors appear. Third, Color Fidelity Refinement (CFR) is a training-free technique that modulates the guidance scale during generation to pull colors toward more realistic ranges.

Input Text → T2I Model → Initial Image
                ↓
            CFM Evaluation
                ↓
         CFR Refinement ← Attention Maps
                ↓
           Final Image

Think of this like a photo editor with a “realism dial.” The CFD is like having a master photographer sort millions of images from “fake-looking” to “camera-authentic.” The CFM learns to see what the photographer sees—subtle color relationships that signal real vs synthetic. The CFR is the actual dial that can turn down the artificial saturation boost that T2I models naturally apply. When CFM detects areas that look too processed, CFR’s attention mechanism knows exactly where to apply more conservative color guidance.

Key Concepts

• Color Fidelity: Not about color accuracy (matching a reference) but about color authenticity—whether the color relationships and saturation levels match what cameras actually capture. A sunset can be orange, but there’s a difference between camera-orange and Instagram-orange.

• Spatial-Temporal Guidance Modulation: During diffusion generation, the guidance scale controls how strongly the model follows the text prompt. CFR adaptively adjusts this scale across different image regions and generation timesteps, reducing guidance where colors are becoming unnaturally vivid.

Framework Shift

Before (mainstream approach):        After (this paper):
Text → Model → Image                Text → Model → Image
  ↓                                   ↓        ↓
Human Rating                        CFM Score  CFR Adjust
  ↓                                   ↓        ↓
"Looks good!"                      Realism    Refined
                                   Check      Output

From optimizing for visual impact to optimizing for photographic authenticity, the core shift is measurement-driven refinement.

Expert Assessment

Problem choice: This addresses a real gap that matters for practical applications. The Instagram-ification of AI-generated “realistic” images is a genuine limitation that affects commercial use cases. The problem sits at the intersection of perceptual psychology and generative modeling—fertile ground.

Method maturity: The approach is methodologically sound but not groundbreaking. CFD is valuable dataset work, CFM is standard multimodal learning, and CFR is a clever but incremental guidance technique. The real contribution is recognizing and formalizing the color fidelity problem rather than inventing new architectures.

Experimental integrity: The 1.3M image dataset with ordered realism levels is substantial. Baselines seem fair, though I’d want to see more analysis of edge cases where “realistic” color might legitimately be vivid (neon signs, tropical scenes). The human evaluation methodology could be stronger—color perception varies significantly across individuals.

Writing quality: The paper clearly articulates the problem and solution. The related work section could better position this against other realism metrics. The technical details of CFR’s attention mechanism deserve more explanation—it’s the most novel component but gets the least space.

Verdict: weak accept — Addresses a real problem with a solid if incremental solution, plus contributes a valuable dataset.

Takeaways

The core insight—that evaluation metrics shape generation quality in subtle ways—transfers broadly. Any domain where “good” and “realistic” diverge (synthetic speech, virtual environments, data augmentation) could benefit from similar authenticity-focused metrics. The spatial-temporal guidance modulation technique could work for other attributes beyond color. Most practically, the CFD dataset provides a template for building ordered realism datasets in other domains.