NoOp-Bench

The official benchmark release accompanying "The Drift Kernel: Why Diffusion Models Change Even When Told Not To" (CVPR 2026).

Even when instructed to "do nothing," diffusion models still change their input. NoOp-Bench is the first large-scale benchmark quantifying this effect — no-op drift — across four diffusion architectures, four visual domains, and a controlled noise-strength sweep, enabling reproducible measurement of the Drift Kernel K_M(σ) introduced in the paper.

---

Dataset Summary

Diffusion image-editing pipelines are commonly assumed to preserve content under a null ("make no change") instruction. NoOp-Bench shows this assumption fails measurably and predictably: identity drift scales quadratically with noise strength, is largely independent of prompt wording, and differs systematically between variance-driven models (SD1.5, SD2.1, SDXL) and instruction-tuned models (InstructPix2Pix).

This release contains everything needed to reproduce every table and figure in the paper:

• 10,000+ source images across 4 visual domains (aerial, faces, natural scenes, artwork)
• 120,000+ baseline generations at fixed strength (σ = 0.3) across 4 diffusion models
• 9,600 ablation samples sweeping σ ∈ {0.1, 0.2, 0.3, 0.4} under null and copy prompts
• Evaluation code for MSE / MAE / PSNR / SSIM / LPIPS / CLIP metrics and drift-kernel fitting
• Precomputed metrics for every reported table, so results can be reproduced without regenerating a single image

Key Results

Baseline drift at σ = 0.3, aggregated over 120,000 comparisons (10,000 images × 3 null prompts × 4 models):

Model	Mean MSE	Std	MAE	PSNR	SSIM	Drift regime
SD1.5	0.0059	±0.0061	0.0452	24.52	0.6976	Variance-driven
SD2.1	0.0069	±0.0064	0.0511	23.27	0.6654	Variance-driven
SDXL	0.0085	±0.0072	0.0579	22.05	0.6304	Variance-driven
InstructPix2Pix	0.0052	±0.0071	0.0426	25.89	0.7802	Edit-driven

Fitted drift-kernel coefficients K_M(σ) ≈ k_Mσ² + c_M (null prompts):

Model	kM	cM	R²
SD1.5	0.0345	0.0028	0.960
SD2.1	0.0685	0.0016	0.979
SDXL	0.0710	0.0024	0.964

Full derivations, per-domain breakdowns, and the prompt-independence analysis (null vs. copy prompts differ by <17%) are in the paper and supplementary/Drift_Kernel_Supplementary.pdf.

Dataset Structure

noop-bench/
├── noopbench_inputs_public.zip        # source images: aerial + faces + natural_scenes (8,530 images)
├── fetch_artwork.py                   # reconstructs the artwork domain locally (see Licensing)
├── output_sd15.zip                    # SD1.5 baseline generations
├── output_sd21.zip                    # SD2.1 baseline generations
├── output_sdxl.zip                    # SDXL baseline generations
├── output_instructpix2pix.zip         # InstructPix2Pix baseline generations
├── ablation_outputs_null_prompt.zip   # strength-sweep ablation, null prompts (4,800 samples)
├── ablation_outputs_copy_prompt.zip   # strength-sweep ablation, copy prompts (4,800 samples)
├── LICENSE                            # per-source licensing terms and attribution
└── supplementary/
    ├── Drift_Kernel_Supplementary.pdf # extended theory, proofs, synthetic decoder validation
    ├── README.txt
    ├── code/                          # metric computation, plotting, kernel fitting
    ├── metrics/                       # precomputed CSVs backing every paper table
    └── benchmark/
        └── ablation_100subset/        # 100-image mini-benchmark (25/domain) for fast iteration

Source domains

Domain	Source dataset	Images	Bundled directly?
aerial	EuroSAT	2,000	✅
natural_scenes	COCO val2017	3,000	✅
faces	FFHQ	3,530	✅
artwork	WikiArt	2,500	Run fetch_artwork.py

All images are center-resized to 512×512 PNG/JPG.

Models evaluated

SD1.5, SD2.1, SDXL, InstructPix2Pix — 15 DDIM steps, guidance scale 5.0, strength σ = 0.3 for baseline generation (σ ∈ {0.1, 0.2, 0.3, 0.4} for the ablation sweep), fixed seed 42. Both null prompts ("make no change," "identity," "do nothing") and strict copy prompts are evaluated; see paper Section 4.2 for exact prompt text.

Reproducing Paper Results

The fastest path — using only the precomputed metrics and the 100-image subset, no image generation required:

python supplementary/code/compute_lpips_clip_metrics_local.py --data supplementary/benchmark/ablation_100subset
python supplementary/code/plot_lpips_vs_strength.py
python supplementary/code/plot_lpips_drift_kernel.py
python supplementary/code/generate_metric_tables.py

supplementary/metrics/kernel_coefficients.csv contains the exact k, c, R² values reported in the paper's drift-kernel table. supplementary/metrics/mse_full.csv, mae_full.csv, psnr_full.csv, and ssim_full.csv back the baseline drift table above.

To reproduce baseline/ablation generations from scratch, unpack noopbench_inputs_public.zip, run fetch_artwork.py for the fourth domain, and run your diffusion pipeline of choice at the settings listed above.

Licensing and Attribution

This is a derived work combining four third-party image sources, each under distinct terms, plus model outputs licensed separately from the source images. Read LICENSE before redistributing or building on this dataset — it documents the verified license for each component:

Component	License basis
aerial (EuroSAT)	MIT / open Copernicus Sentinel data — freely redistributable
natural_scenes (COCO)	CC BY 4.0 annotations; images subject to original Flickr per-photo terms
faces (FFHQ)	CC-licensed per-photo; redistributable for non-commercial research with attribution
artwork (WikiArt)	WikiArt's terms prohibit third-party redistribution — not bundled; use fetch_artwork.py
Generated model outputs	CreativeML Open RAIL-M (SD1.5/2.1/XL)

Citation

If you use NoOp-Bench in your research, please cite:

@InProceedings{Ram_2026_CVPR,
    author    = {Ram, Gokul Srinath Seetha and Elavazhagan, Rashmi},
    title     = {The Drift Kernel: Why Diffusion Models Change Even When Told Not To},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {43281-43289}
}

Contact

Gokul Srinath Seetha Ram — s.gokulsrinath@gmail.com Rashmi Elavazhagan — rashmie30@gmail.com