ManLuML/onmanifold-tfg-bird-benchmark

On-Manifold TFG — Fine-Grained Bird Benchmark

A parent–child benchmark for evaluating training-free guidance (TFG) of diffusion models, introduced in Not All Prediction Targets Keep Training-Free Diffusion Guidance on the Manifold (ECCV 2026).

The task it defines: classifier-free guidance steers generation toward a coarse ImageNet parent class (e.g. goldfinch), while training-free guidance steers toward a fine-grained child species (e.g. AMERICAN GOLDFINCH). Measuring how far a guided sample drifts from the child distribution is what the paper's Child FID (C-FID) metric captures.

What is ours, and what is not

The images are not ours. They come from Piosenka's public BIRDS 525 SPECIES dataset (CC0). What this repository adds is the benchmark layer:

• the 143-species → 30-ImageNet-parent hierarchy (only ImageNet classes with ≥2 fine-grained children are kept; 2–20 children per parent, mean 4.8),
• per-sample imagenet_parent / imagenet_parent_name annotations,
• a pinned snapshot of the source data.

Configurations

Config	Images	Classes	train / validation / test
full	89,885	525	84,635 / 2,625 / 2,625
benchmark	24,439	143	23,009 / 715 / 715

benchmark is the benchmark proper; full is the complete source snapshot it was drawn from. In full, the 382 species outside the benchmark carry imagenet_parent = -1 and an empty imagenet_parent_name.

Labels are the canonical alphabetical ordering of the 525 species — the same ordering used by the guide and evaluation classifiers below. Indices are identical across both configs (the ClassLabel feature carries all 525 names either way), so an index means the same species in each, and matches finegrained_bird_mapping.json shipped alongside this dataset.

Usage

from datasets import load_dataset

bench = load_dataset("ManLuML/onmanifold-tfg-bird-benchmark", "benchmark", split="train")
ex = bench[0]
ex["image"]                 # PIL.Image, RGB (224x224 except where noted below)
ex["label_name"]            # e.g. "AMERICAN GOLDFINCH"
ex["imagenet_parent"]       # e.g. 11
ex["imagenet_parent_name"]  # e.g. "goldfinch"

Evaluation protocol used in the paper

Generation is at 256×256, 64 samples per species across all 143 species (9,152 images per condition), with the guidance strength ρ swept per model. Two deliberately different classifiers are used so that validity is not scored by the model that produced the gradient:

• guide (gradient source): dennisjooo/Birds-Classifier-EfficientNetB2
• eval (validity): chriamue/bird-species-classifier

Provenance

• Upstream: chriamue/bird-species-dataset, a mirror of Kaggle: BIRDS 525 SPECIES by Gerald Piosenka.
• Snapshot taken 2026-07-28; SHA-256 of every source archive is recorded in the repository's SHA256SUMS.txt.
• Images are RGB JPEG, collected by species-name search, deduplicated, and cropped so the bird occupies at least 50% of pixels. We confirmed there are no duplicate files and no train/validation/test leakage (byte-level hash over all 89,885 files).

Resolution is not uniform

211 of 89,885 images (0.23%) are not 224×224 (92×225 up to 897×575), and 210 of them are the single species PLUSH CRESTED JAY, which is part of the benchmark. Everything else is exactly 224×224. Resize explicitly rather than assuming a fixed tensor shape.

Licence

CC0-1.0, inherited from 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