Rfid Rfvd Psnr

---
name: rfid-rfvd-psnr
description: Measures the reconstruction fidelity of the visual tokenizer by comparing discrete latent reconstructions to original images/videos. Use when the user has predictions and gold and needs to compute rFID.
metadata:
  skill_kind: metric
  source_arxiv: 2604.13030
  bibtex_key: han2026generativerefinement
  confidence: high
---

# rfid-rfvd-psnr

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

## What this evaluates

Measures the reconstruction fidelity of the visual tokenizer by comparing discrete latent reconstructions to original images/videos.

## Datasets

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

## Metrics

- `rFID` **(primary)** — range: other
  - Reconstruction Fréchet Inception Distance. Lower indicates better perceptual fidelity of reconstructed images.

## Input / output format

**Input**: Original image

**Output**: Reconstructed image from discrete tokens

## Scoring recipe

```python
recon_images = []
for img in test_set:
    tokens = tokenizer.encode(img)
    recon = tokenizer.decode(tokens)
    recon_images.append(recon)
rfid = compute_fid(imgs, recon_images)
```

## Common pitfalls

- rFID is computed on reconstructed images, not generated ones.
- Metrics like PSNR and SSIM measure pixel-level fidelity, while rFID/LPIPS measure perceptual quality.

## Evidence (verbatim from paper)

> our tokenizer demonstrates state-of-the-art reconstruction performance on the 256x256 ImageNet benchmark. Utilizing four HBQ rounds, it achieves a remarkable rFID of 0.56.

## 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