With the growing demand for real-time video enhancement in live applications, existing methods often struggle to balance speed and effective exposure control, particularly under uneven lighting. We introduce RRNet (Rendering Relighting Network), a lightweight and configurable framework that achieves a state-of-the-art tradeoff between visual quality and efficiency. By estimating parameters for a minimal set of virtual light sources, RRNet enables localized relighting through a depth-aware rendering module without requiring pixel-aligned training data. This object-aware formulation preserves facial identity and supports real-time, high-resolution performance using a streamlined encoder and lightweight prediction head. To facilitate training, we propose a generative AI-based dataset creation pipeline that synthesizes diverse lighting conditions at low cost. With its interpretable lighting control and efficient architecture, RRNet is well suited for practical applications such as video conferencing, AR-based portrait enhancement, and mobile photography. Experiments show that RRNet consistently achieves the best overall result in low-light enhancement, localized illumination adjustment, and glare removal.
Figures & tables
Fig. 1: NIQE–runtime comparison on 1080p VCD videos using an RTX 3090 (lower is better). RRNet ( N -frame) estimates lighting every N frames and reuses it between updates, amortizing the estimation cost. For N=1,3,10 , runtime is 17.0, 6.3, and 2.6 ms per frame, respectively.
Fig. 2: VisualComparisons of various low-light enhancement methods on test images from the FFHQ, VV, and VCD datasets show that RRNet achieves superior exposure balance and preserves skin tones more effectively than other state-of-the-art methods.
Fig. 3: Architecture of RRNet . (a) Overall framework with Lighting Parameter Regression Module (LPRM) , Rendering Module (RM) , and optional Albedo Generation Module (AGM) ; the dashed line in LPRM indicates shared weights between encoders E0 and E1 . (b) RM details; M , N , and θ∗ denote depth, surface normal, and estimated lighting parameters.
Fig. 4: FFHQL data generation pipeline. (Top) Abstract flow: an FFHQ portrait is fed to an extended IC-Light [ 30 ] diffusion conditioned on sampled lighting controls (direction, pattern, intensity, color temperature) to produce N identity-consistent variants; an expert vote then selects the most uniformly lit and aesthetically pleasing one as the GT. (Bottom) Visualization on two FFHQ identities: each source portrait yields N variants (boxed image is the voted GT); each non-GT variant is paired with the same voted GT to form multiple training pairs (two such pairs shown for the first identity).
Method
NPE
LIME
MEF
VV
DICM
VCD
Avg.
Portrait
EnlightenGAN
4.11
3.72
3.32
2.68
3.45
4.74
3.67
2.63
DeepUPE
3.67
4.14
3.68
3.22
3.89
5.13
3.96
3.14
RetinexNet
4.46
4.42
3.98
2.98
4.20
3.68
3.95
2.95
PairLIE
4.02
4.51
4.16
3.66
4.09
4.56
4.08
3.53
RetinexMamba
3.55
3.88
3.32
5.62
3.57
5.28
4.20
5.57
CWNet
3.65
4.46
4.38
2.62
3.83
5.12
4.01
2.64
TABLE I: NIQE comparison on various datasets. Lower NIQE scores indicate higher quality. Throughout all tables in this paper, bold marks the best result and underline marks the second-best. Methods listed in the lower half of the table are expected to achieve real-time performance during testing. Rows marked † are values reported by the original publications; “—” indicates a benchmark not reported.
Modification
NIQE ↓
FID ↓
Time (ms) ↓
None (Baseline)
3.69
22.84
17.0
Single-branch
4.00
23.35
9.5
Remove AGM
4.05
23.03
15.7
Single-branch + Remove AGM
4.06
23.51
8.9
TABLE II: Ablation study on model architecture (FFHQL dataset). Baseline: dual-branch LPRM with AGM and 9 virtual lights. Time is measured on an NVIDIA RTX 3090 GPU.
# Lights
NIQE ↓
FID ↓
Time (ms) ↓
3
3.78
23.71
16.5
6
3.71
22.97
16.7
9
3.69
22.84
17.0
12
3.72
24.25
17.5
TABLE III: Ablation study on number of virtual lights (FFHQL validation set).
Loss Configuration
NIQE ↓
FID ↓
Lpixel only
3.78
23.01
Lpixel+Lroi
3.72
24.69
Lpixel+Lroi+λrLreg
3.71
23.05
Full ( +λretLretinex )
3.69
22.84
TABLE IV: Ablation study on loss functions (FFHQL dataset).
Fig. 5: Comparison of frame enhancement on an unbalanced exposure video from the VCD dataset. From top to bottom: input frames and results of ZeroDCE++, SNRNet, StableLLVE, and our method.
Method
Time (ms) ↓
T-Warp ↓
Flicker ↓
Input (reference)
–
1.05
1.00
Zero-DCE++
1.2
1.37
1.46
SNRNet
29.0
1.58
1.71
FastLLVE
7.4
1.77
2.17
StableLLVE
2.5
1.11
1.12
RRNet (1-frame, default)
17.0
1.36
1.35
TABLE V: Temporal consistency on VCD video clips. T-Warp is the RAFT-based flow-warping error on output frames ( ×10−2 ); Flicker is the ratio of output to input warp error ( 1.00 means no added flicker beyond natural pixel motion). RRNet additionally exposes a configurable lighting-update interval at β=0.95 .
Fig. 6: Qualitative glare removal: with AGM enabled, RRNet’s renderer operates on the estimated albedo and locally suppresses strong directional sources while preserving facial structure that pixel-wise baselines either dim or amplify.
Department of Computer Science and Engineering, BRAC University, Dhaka, 1205, Bangladesh · Department of Computer Science and Engineering, Bangladesh University of Engineering and Technology, Dhaka, 1205, Bangladesh