bmltera/color-hybrid-illusions

Color Hybrid Illusions Dataset

A benchmark dataset of 177 image pairs for studying how vision-language models (VLMs) resolve conflicting visual cues. Each image depicts one entity in color and a different entity in grayscale, created using Factorized Diffusion.

Overview

When you view a color hybrid image in full color, you see one object (e.g., a bird). When you convert it to grayscale, a different object emerges (e.g., a flower). This dataset uses that conflict to test whether VLMs rely more on chromatic (color) or luminance (grayscale/shape) cues for object recognition.

Key finding: Across 11 VLMs and 3,894 predictions, most models exhibit grayscale bias (avg gray accuracy 0.681 vs. color accuracy 0.554), suggesting VLMs generally privilege shape and luminance structure over color information.

How the Dataset Was Generated

Images were generated using Factorized Diffusion (Geng et al., ECCV 2024), which decomposes a diffusion model's denoising process into separate linear components — in this case, grayscale (luminance) and color (chrominance) channels. Each component is conditioned on a different text prompt during sampling, producing a single image that depicts one object in color and a different object in grayscale structure.

The underlying diffusion model is DeepFloyd IF, a pixel-based cascaded diffusion pipeline that generates 1024×1024 images. Text prompts are encoded with a T5 text encoder and guide the denoising process across both views.

Pipeline:

1. Prompt pairing — Each image pair is generated from two prompts: one describing a grayscale object (e.g., "a shaded sketch of a lily") and one describing a color object (e.g., "a vivid poster of a finch").
2. Factorized sampling — The diffusion model denoises both the grayscale and color components simultaneously, each conditioned on its respective prompt.
3. Human auditing — From an initial pool of 2,400 generated pairs, each image was manually reviewed and assigned a quality tier. Only pairs that successfully produced a visible illusion were retained, resulting in the final set of 177 pairs.

Dataset Structure

• dataset.json — Metadata for all 177 pairs, including prompts, object labels, and quality tiers.
• images/ — 354 PNG images (one color c + one grayscale g per pair).

Naming Convention

Images are named {number}c.png (color view) and {number}g.png (grayscale view), zero-padded to 4 digits. For example, pair #98 → 0098c.png and 0098g.png.

Metadata Fields

Field	Description
number	Image pair ID
greyscale	Prompt used for the grayscale component
color	Prompt used for the color component
quality	Human-rated quality tier: L (low), M (medium), H (high)
grey_object	Ground-truth object label for the grayscale view
color_object	Ground-truth object label for the color view

Quality Tiers

Quality tiers assess how well each generated illusion decouples luminance structure from chromatic information:

• H (High): Clear, drastic difference between entities across views — the illusion is immediately apparent
• M (Medium): Moderate distinction between entities
• L (Low): Less distinction; both views may partially resemble each other

Example

Grayscale View → "flower"	Color View → "bird"
0098g.png	0098c.png

Benchmark Results

Per-Model Performance (Forced-Choice)

Model	Overall Acc.	Gray Acc.	Color Acc.	Δ	Bias
ALIGN	0.701	0.785	0.616	+0.169	Gray
SigLIP	0.684	0.746	0.621	+0.124	Gray
LLaVA-1.6	0.667	0.802	0.531	+0.271	Gray
SmolVLM	0.655	0.729	0.582	+0.147	Gray
Qwen2-VL	0.653	0.695	0.610	+0.085	Gray
GPT-4o-mini	0.644	0.689	0.599	+0.090	Gray
LLaVA-1.5	0.633	0.757	0.508	+0.249	Gray
CLIP	0.630	0.802	0.458	+0.345	Gray
GPT-5.5	0.540	0.497	0.584	−0.087	Color
BLIP-2	0.500	0.435	0.565	−0.130	Color
Moondream2	0.483	0.548	0.418	+0.130	Gray

Architecture Families

Family	Models	Avg. Accuracy
Contrastive	CLIP, ALIGN, SigLIP	0.671
Generative (Q-Former)	BLIP-2	0.500
Instruction-tuned LLM	LLaVA-1.5, LLaVA-1.6, Qwen2-VL	0.651
Compact VLM	SmolVLM, Moondream2	0.569
Proprietary API	GPT-4o-mini, GPT-5.5	0.592

Intended Use

This dataset is intended for:

• Evaluating VLM cue arbitration — testing whether models rely on shape/luminance or color when the two conflict
• Benchmarking multimodal robustness — assessing model performance on perceptually ambiguous inputs
• Studying representation bias — understanding how training objectives (contrastive, generative, instruction-tuned) influence visual feature weighting

Citation

@misc{li2026entityrecognition,
  title={Entity Recognition with Vision Language Models on Diffusion-Based Color Hybrid Illusions},
  author={Bill Li and Paul Junver Soriano and Rahul Koonantavida},
  year={2026},
  institution={San Jos\'{e} State University}
}

Links

• Project Website: hybrid-color-images.vercel.app
• Factorized Diffusion Paper: Geng et al., ECCV 2024
• Visual Anagrams Paper: Geng et al., CVPR 2024