Imagenet C2i Fid Is

---
name: imagenet-c2i-fid-is
description: Evaluates class-conditional image generation fidelity and diversity on ImageNet 256x256. It measures how closely the distribution of generated images matches real images and how well the model covers all classes. Use when the user wants to benchmark on ImageNet, or asks about evaluating this task. Reports FID.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.13030
  bibtex_key: han2026generativerefinement
  confidence: high
---

# imagenet-c2i-fid-is

> Generative Refinement Networks for Visual Synthesis — Jian Han et al. (arXiv:2604.13030, 2026)

## What this evaluates

Evaluates class-conditional image generation fidelity and diversity on ImageNet 256x256. It measures how closely the distribution of generated images matches real images and how well the model covers all classes.

## Datasets

- **ImageNet** — total ?; splits: test (50000)

## Metrics

- `FID` **(primary)** — range: other
  - Fréchet Inception Distance computed over 50K generated samples against 50K validation samples. Lower is better.
- `IS` — range: other
  - Inception Score computed over 50K generated samples. Higher is better.

## Input / output format

**Input**: Integer class label (0-999)

**Output**: 256x256 RGB image

## Scoring recipe

```python
images = []
for class_id in range(1000):
    for _ in range(50):
        images.append(model.generate(class_id))
fid = compute_fid(images, imagenet_val_50k)
is_score = compute_is(images)
```

## Common pitfalls

- FID and IS are evaluated on exactly 50K samples, not the full validation set.
- Inference requires grid-searching Classifier-Free Guidance (CFG) scale and interval for optimal results.

## Evidence (verbatim from paper)

> FID *[fid]* and IS *[inception_score]* of 50K samples are evaluated.

## Citation

```bibtex
@misc{han2026generativerefinement,
  title={Generative Refinement Networks for Visual Synthesis},
  author={Jian Han et al.},
  year={2026},
  note={arXiv:2604.13030}
}
```

- arXiv: 2604.13030