VybBench 1.0

Almost isn't good enough. VybBench measures how well models maintain design fidelity by comparing how effectively they generate specific fonts and colors, and perform non-destructive edits.

Our dataset consists of 145 fonts, 100 colors, and 200 edits. We have each model perform each task three times and score each generation individually.

We compare top image generation models: ChatGPT Image (gpt-image-2.5-sunburst), Meta Muse, and Recraft v4.1. We also report the results for Vyb 1.0.

Full Disclosure: VybBench was built by Coco Labs, the makers of Vyb. The full methodology and rationale are in the launch post.

Results

Font fidelity

We sample 145 Google Fonts. For every font we choose a word to generate. We generate that word programmatically as well as through model calls. We compute the IoU of the ink between the ground truth generation and each model generation. The higher the score the better. As each model is allowed to scale and place the text individually, we compute the noise floor by measuring the IoU of multiple ground truth generations with the same font.

Model	Mean	Worst 5% of cases
Vyb 1.0	0.90	0.78
ChatGPT Image (gpt-image-2.5-sunburst)	0.76	0.28
Muse (meta/muse-image)	0.51	0.18
Recraft (v4.1)	0.46	0.14
Measurement noise floor	0.94	0.88

An example from the benchmark. The ground truth is shown as is each model's output. Only Vyb generates an output that is close to the ground truth.

Color fidelity

To measure color fidelity, we pick 100 different colors. We ask each model to generate a rectangle in the specified color, given as a hex code. We then calculate the per-channel error for each of R, G, and B (out of 255) and report the maximum across channels.

Model	Mean	Worst 5% of cases
Vyb 1.0	0.0	0.0
ChatGPT Image (gpt-image-2.5-sunburst)	3.9	8
Muse (meta/muse-image)	5.7	10
Recraft (v4.1)	27.2	56

Here we see that only Vyb reproduces every requested color exactly.

Edit fidelity

We provide every model the same starting design and the same edit prompt. For Vyb, the design is provided in the layered SVG format it accepts. For the leading image generation models, the SVG is rendered and provided in the PNG format they accept.

For each model generation, we measure the fraction of pixels in the image that were changed and have an LLM judge score whether the edit was performed (acceptance rate).

We report the fraction of pixels changed (lower is better) and acceptance rate (higher is better).

Model	Pixels changed, mean	Worst 5% of cases	Acceptance rate
Vyb 1.0	7.3%	31.4%	99.2%
ChatGPT Image (gpt-image-2.5-sunburst)	95.5%	99.7%	100%
Muse (meta/muse-image)	91.9%	99.6%	98.0%
Recraft (v4.1)	98.9%	100%	51.2%

An example of an edit where the request is to change the color of the "30% OFF" text to coral pink. Models other than Vyb change unrelated parts of the image.

Reproduce

To judge new outputs, you need a Gemini API key.

pip install google-genai
GEMINI_API_KEY=... python -m scoring.judge

You can rescore by running:

pip install numpy pillow scikit-image
python -m scoring

All designs are synthetic. They contain fictional brands with AI-generated imagery.

Run: October 2026.

Add your model

We are happy to include your model. Create a discussion on this dataset with your model's outputs (follow the outputs/ layout, provide three attempts per case).

Citation

@misc{vybbench2026,
  title = {VybBench 1.0: A Design Fidelity Benchmark},
  author = {{Coco Labs}},
  year = {2026},
  url = {https://huggingface.co/datasets/cocolabs/vybbench}
}