Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 36 of the 38 references it cites were found in a registry or an index; 7 of the 20 sources its claims cite were checked for support.

Temporally consistent low-light face video enhancement via video-to-video conditional diffusion

Open access
Ding, Xiaofeng, He, Kailin, Sun, Huo, Yang, Juying · 2026 · Scientific Reports

Abstract

Low-light face videos suffer from severe noise and detail loss, limiting their use in surveillance and photography applications. To address these challenges, this paper proposes DL-Diff, a novel low-light face video enhancement framework that formulates this task as a conditional video-to-video (V2V) generation problem based on pre-trained Latent Diffusion Models (LDMs). DL-Diff extends pre-trained text-to-video models through three components: a pseudo-3D UNet backbone, a restoration component for spatial detail recovery, and a temporal component for inter-frame consistency. A multi-stage training strategy enables efficient domain adaptation from images to videos. Experiments on DID and SDSD datasets demonstrate that DL-Diff achieves superior performance in both perceptual quality (FID: 41.29, LPIPS: 0.17) and temporal consistency (AB(Var): 25.40, MABD: 0.08), significantly outperforming existing methods. The framework produces high-quality videos with realistic visual effects and no flickering artifacts, particularly excelling in extremely dark scenarios. This work demonstrates the potential of leveraging pre-trained diffusion models for video enhancement tasks.

Overview

CoChat Summary

TL;DR

DL-Diff reformulates low-light face video enhancement as conditional video-to-video generation using a pre-trained latent diffusion model, adding a restoration component and a temporal component to a pseudo-3D UNet. It reports strong perceptual quality and temporal consistency on DID and SDSD, with higher inference cost than feed-forward baselines.

Key findings

  • On DID, DL-Diff reports FID 41.29, LPIPS 0.17, PSNR 22.92 dB, SSIM 0.72, AB(Var) 25.40, and MABD 0.08.
  • Ablation: full multi-stage training raises PSNR from 18.96 to 22.92 dB and SSIM from 0.44 to 0.72; removing the temporal component worsens AB(Var) to 27.77 and MABD to 0.10, recovering to 25.40 and 0.08 when restored.
  • On DID, DL-Diff ranks first in both temporal metrics, with AB(Var) 9.06 lower than MBLLEVEN and MABD 0.03 lower than StableLIVE; on SDSD it is second-best, behind the SDSD-trained method.
  • Qualitative results claim better color restoration and detail in extremely dark scenes, with no flickering artifacts, but inference takes about 45 seconds per 16-frame 256×256 clip on an NVIDIA A100 with 100 DDIM steps versus 0.5–1.2 seconds for traditional methods.

Methods

  • DL-Diff builds on frozen Stable Diffusion v1.5 (~860M parameters), extends its 2D UNet to a pseudo-3D UNet, and trains a ControlNet-like restoration component (~370M parameters) plus an AnimateDiff-like temporal component (~85M parameters), totaling ~455M trainable parameters.
  • Training uses a three-stage pipeline: Stage 1 trains the restoration component on randomly sampled low-light/normal-light image pairs; Stage 2 trains the temporal component on normal-light 16-frame video clips; Stage 3 jointly fine-tunes both components on paired low-light/normal-light videos. Training is at 256×256 resolution with T=1000 diffusion steps and 100 DDIM inference steps.
  • Evaluation uses DID (413 aligned low-light/normal-light video pairs, up to 2560×1440) and SDSD (150 video pairs, 70 indoor and 80 outdoor, 100–300 frames at 1920×1080, non-aligned), with FID, LPIPS, PSNR, SSIM, AB(Var), and MABD metrics plus ablations.

Limitations

  • Stated: diffusion inference cost remains higher than traditional methods; PSNR does not consistently surpass all baselines, reflecting a trade-off between perceptual quality and pixel-level reconstruction; the framework focuses on face videos, and extending to complex general scenes would require additional architectural considerations.
  • Evident from the text: inference is about 45 seconds per 16-frame 256×256 clip on an A100 with 100 DDIM steps, so it is not real-time; evaluation is limited to DID and SDSD face-video settings.
  • Evident from reported comparisons: some baselines exceed DL-Diff on pixel-level metrics on DID, e.g., MBLLEVEN reports PSNR 24.82 dB and FlowIE reports SSIM 0.92, while DL-Diff reports PSNR 22.92 dB and SSIM 0.72.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41598-026-44219-8

Cites

36 of 38 references verified2 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 The proposed dark-to-light diffusion (DL-Diff) framework for low-light face video enhancement. It utilizes a pseudo-3D UNet based on pre-trained LDMs, incorporating a Restoration Component for spatial detail enhancement and a Temporal Component for inter-frame consistency. The model is trained via a multi-stage training process to adapt pre-trained T2V models for V2V tasks.
Fig. 1 The proposed dark-to-light diffusion (DL-Diff) framework for low-light face video enhancement. It utilizes a pseudo-3D UNet based on pre-trained LDMs, incorporating a Restoration Component for spatial detail enhancement and a Temporal Component for inter-frame consistency. The model is trained via a multi-stage training process to adapt pre-trained T2V models for V2V tasks.CC BY-NC-ND · PubMed Central
Algorithm 1 Multi-stage training strategy for DL-Diff.
Algorithm 1 Multi-stage training strategy for DL-Diff.CC BY-NC-ND · PubMed Central
Fig. 2 Qualitative comparison of LLFVE methods on DID dataset. Visual comparison of different low-light face video enhancement methods: (a) dark input, (b) MBLLEVEN, (c) StableLIVE, (d) SMID, (e) DL-Diff, and (f) ground truth. DL-Diff demonstrates superior performance in color restoration and detail preservation.
Fig. 2 Qualitative comparison of LLFVE methods on DID dataset. Visual comparison of different low-light face video enhancement methods: (a) dark input, (b) MBLLEVEN, (c) StableLIVE, (d) SMID, (e) DL-Diff, and (f) ground truth. DL-Diff demonstrates superior performance in color restoration and detail preservation.CC BY-NC-ND · PubMed Central
Fig. 3 Module insertion points for ablation study. Left: encoder (input \rightarrow Down1–Down4 \rightarrow middle). Center: decoder (middle \rightarrow Up1–Up4 \rightarrow output). Right: restoration branch (low-light video \rightarrow restoration encoder). At each decoder block, the input is formed by adding the corresponding encoder skip to the restoration residual r_i. Temporal attention modules (inline) are inserted after Down2, Down4, Middle, and Up2.
Fig. 3 Module insertion points for ablation study. Left: encoder (input \rightarrow Down1–Down4 \rightarrow middle). Center: decoder (middle \rightarrow Up1–Up4 \rightarrow output). Right: restoration branch (low-light video \rightarrow restoration encoder). At each decoder block, the input is formed by adding the corresponding encoder skip to the restoration residual r_i. Temporal attention modules (inline) are inserted after Down2, Down4, Middle, and Up2.CC BY-NC-ND · PubMed Central
Fig. 4 Training stability across the three-stage pipeline. Loss curves for Stage 1 (restoration component on image pairs), Stage 2 (temporal component on normal-light videos), and Stage 3 (joint fine-tuning on paired videos), demonstrating stable convergence.
Fig. 4 Training stability across the three-stage pipeline. Loss curves for Stage 1 (restoration component on image pairs), Stage 2 (temporal component on normal-light videos), and Stage 3 (joint fine-tuning on paired videos), demonstrating stable convergence.CC BY-NC-ND · PubMed Central
Fig. 5 Qualitative comparison of LLFVE methods on DID dataset. Visual comparison of different low-light face video enhancement methods: (a) dark input, (b) MBLLEVEN, (c) StableLIVE, (d) SMID, (e) DL-Diff, and (f) ground truth. DL-Diff demonstrates superior performance in color restoration and detail preservation.
Fig. 5 Qualitative comparison of LLFVE methods on DID dataset. Visual comparison of different low-light face video enhancement methods: (a) dark input, (b) MBLLEVEN, (c) StableLIVE, (d) SMID, (e) DL-Diff, and (f) ground truth. DL-Diff demonstrates superior performance in color restoration and detail preservation.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

2 Valid · 2 Not valid · 4 Not checked · 78 Own results and interpretations

32 more claims

Create a free account to see all 86 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 42
  • Probably not a substantial claim. 10

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 36 references

Research works printed but not linked: 1

  1. Dosovitskiy, A. & Brox, T. Generating images with perceptual similarity metrics based on deep networks. Adv. Neural Inf. Process. Syst. 29 (2016).

Metadata

Authors
Ding, X., He, K., Sun, H., & Yang, J.
Affiliations
  • Sichuan University Jinjiang College (CN)
  • Sichuan Railway College (CN)
Corresponding author
Ding, Xiaofeng
Venue
Scientific Reports
Published
2026-03-18
Volume
16
Issue
1
ISSN
2045-2322 (ISSN-L)
OpenAlex
W7139019837
Cited by
0
References
36
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Enhancement Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image Processing Techniques
Keywords
latent diffusion models, temporal coherence, text-to-video diffusion models

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Temporally consistent low-light face video enhancement via video-to-video conditional diffusion. https://app.cochat.ai/papers/10.1038/s41598-026-44219-8
The paper
Ding, X., He, K., Sun, H., & Yang, J. (2026). Temporally consistent low-light face video enhancement via video-to-video conditional diffusion. Scientific Reports. https://doi.org/10.1038/s41598-026-44219-8

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.