ManLuML/onmanifold-tfg-butterfly-benchmark

On-Manifold TFG — Fine-Grained Butterfly Benchmark

The second fine-grained domain from Not All Prediction Targets Keep Training-Free Diffusion Guidance on the Manifold (ECCV 2026), used to show that the paper's prediction-target hierarchy is not an artifact of a single visual domain.

Same structure as the bird benchmark: classifier-free guidance targets a coarse ImageNet butterfly parent, training-free guidance targets a fine-grained species, and Child FID measures the drift away from the species distribution.

What is ours, and what is not

The images come from Piosenka's public Butterfly & Moths Image Classification 100 species dataset (CC0). What this repository adds:

• the 34-species → 6-ImageNet-parent hierarchy (admiral, ringlet, monarch, cabbage_butterfly, sulphur_butterfly, lycaenid_butterfly),
• per-sample imagenet_parent / imagenet_parent_name annotations,
• a pinned snapshot of upstream version 13, the data the paper used.

The 34-species subset is a taxonomic pruning of the upstream 100 classes. Moths (~20 classes), swallowtails, large tropicals, and skippers are dropped because none of them has a clean ImageNet parent. The resulting per-parent branching is admiral 8 / ringlet 2 / monarch 6 / cabbage_butterfly 6 / sulphur_butterfly 4 / lycaenid_butterfly 8 — an average of 5.7 children per parent, deliberately close to the bird benchmark's 4.8 so the two domains are comparable.

Configurations

Config	Images	Classes	train / validation / test
full	13,594	100	12,594 / 500 / 500
benchmark	4,781	34	4,441 / 170 / 170

In full, the 66 species outside the benchmark carry imagenet_parent = -1 and an empty imagenet_parent_name.

Labels are the canonical alphabetical ordering of the 100 species (ADONIS … ZEBRA LONG WING) — the same ordering used by the guide and evaluation classifiers below. Indices are shared across both configs and match finegrained_butterfly_mapping.json shipped alongside this dataset.

Usage

from datasets import load_dataset

bench = load_dataset("ManLuML/onmanifold-tfg-butterfly-benchmark", "benchmark", split="train")
ex = bench[0]
ex["image"]                 # PIL.Image, 224x224 RGB
ex["label_name"]            # e.g. "RED ADMIRAL"
ex["imagenet_parent"]       # e.g. 321
ex["imagenet_parent_name"]  # e.g. "admiral"

Evaluation protocol used in the paper

256 images are generated per species across all 34 species (8,704 images per condition), a budget close to the bird benchmark's 9,152. Guide and eval classifiers differ in architecture as well as weights, so validity is not scored by the gradient source:

• guide (gradient source): a ConvNeXt-base fine-tuned by us on the upstream 100-species split (CNN)
• eval (validity): dima806/100_butterfly_types_image_detection (ViT)

Provenance

• Upstream: Kaggle: Butterfly & Moths Image Classification 100 species by Gerald Piosenka, version 13 (pinned).
• Snapshot taken 2026-07-28; SHA-256 of the source archive is recorded in the repository's SHA256SUMS.txt.
• Images are 224×224 RGB JPEG.
• Excludes the pretrained EfficientNetB0 .h5 classifier bundled in the upstream archive; the images and labels are otherwise complete.

Licence

CC0-1.0, as declared on the source dataset. Attribution is not legally required, but please credit Gerald Piosenka for the underlying images.

Citation

@inproceedings{lee2026onmanifold,
  title     = {Not All Prediction Targets Keep Training-Free Diffusion Guidance on the Manifold},
  author    = {Lee, Yunsung and Lee, Hyeongmin},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Code: https://github.com/ManLuML/on-manifold-tfg