Lumivid Eval

lumivid-eval

HDR Video Generation via Latent Alignment with Logarithmic Encoding — Korem et al. (2026) (arXiv:2604.11788, 2026)

What this evaluates

Evaluates a model's ability to generate physically plausible, temporally coherent HDR video from standard dynamic range (SDR) inputs. It probes reconstruction fidelity in perceptually uniform HDR spaces, temporal stability across frames, and the model's capacity to recover clipped radiance details using learned visual priors.

Datasets

• ARRI Cinema Footage — total 48; splits: test (48)
• UPIQ — total 30; splits: test (30)

Metrics

• PU21-PSNR (primary) — range: dB
  • Peak Signal-to-Noise Ratio computed after encoding both prediction and ground truth into the PU21 perceptually uniform HDR color space. Reported in decibels (dB).
• ColorVideoVDP JOD — range: [0, 10]
  • Perceptual quality score derived from the ColorVideoVDP visual difference predictor model. Outputs a 0–10 scale where higher values indicate better perceptual fidelity.
• F2F-PSNR — range: dB
  • Frame-to-frame PSNR computed between consecutive frames in the generated video sequence to measure local temporal consistency.
• Flicker — range: percent
  • Temporal stability metric calculated as the standard deviation of per-frame mean luminance, normalized by the overall mean luminance across the sequence.
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity score measuring perceptual distance between generated and ground truth frames.

Input / output format

Input: 8-bit SDR video frames (generated by tone-mapping 12-bit HDR ground truth) or static images converted into 9-frame video sequences.

Output: HDR video frames with scene-linear radiance values (12-bit/16-bit precision).

Scoring recipe

def compute_metrics(pred, gt):
    # Encode to PU21 space for HDR-aware fidelity metrics
    pred_pu21 = encode_pu21(pred)
    gt_pu21 = encode_pu21(gt)
    pu21_psnr = 10 * math.log10(MAX_VAL**2 / mse(pred_pu21, gt_pu21))
    
    # Perceptual & temporal metrics
    jod = colorvideovdp_jod(pred, gt)
    lpips = perceptual_similarity(pred, gt)
    
    # Video-specific temporal stability
    frame_means = np.mean(pred, axis=(1, 2, 3))
    flicker = np.std(frame_means) / np.mean(frame_means)
    f2f_psnr = psnr(pred[1:], pred[:-1])
    
    return pu21_psnr, jod, lpips, flicker, f2f_psnr

Common pitfalls

• Using standard linear PSNR/SSIM instead of the perceptually uniform PU21-PSNR/SSIM, which severely misrepresents HDR reconstruction quality.
• Evaluating video baselines strictly per-frame without computing temporal metrics (Flicker, F2F-PSNR), leading to inflated quality scores despite visible frame-to-frame flicker.
• Confusing the 8-bit SDR input (tone-mapped from 12-bit GT) with the actual ground truth radiance values during preprocessing or metric computation.

Evidence (verbatim from paper)

Fidelity is measured by comparison to the gorund truth hdr iamges and videos uisng via PU21-PSNR and PU21-SSIM*[21]* and ColorVideoVDP JOD*[22]* a 0–10 perceptual scale, that can be applied on both images and videos. Temporal stability is assessed only on the video benchmark, using standard deviation of per-frame mean luminance, normalized by the overall mean (Flicker) and frame-to-frame (F2F) PSNR.

Citation

@misc{korem2026hdrvideo,
  title={HDR Video Generation via Latent Alignment with Logarithmic Encoding},
  author={Korem et al. (2026)},
  year={2026},
  note={arXiv:2604.11788}
}

• arXiv: 2604.11788