How Tencent’s TNC Neural Codec Won 2025 Image & Video Compression Challenges

Competition Results

In December 2025 the VCIP 2025 Low‑Complexity End‑to‑End Image Compression Challenge and the PCS 2025 High‑Compression Intelligent Image Compression Challenge were concluded. Tencent Shannon Lab’s Tencent Neural Codec (TNC) achieved the best visual quality at the same bitrate in both contests, winning the image track (average 0.51 bpp, +1.66 dB PSNR over BPG, up to +2.81 dB) and the video track (subjective MOS +1.20 over VTM, +0.57 over ECM, objective PSNR +3.07 dB over VTM).

TNC Image Codec Architecture

VAE‑INR Hybrid Encoder

TNC introduces a controllable frame‑level and block‑level rate‑distortion optimizer that selects either a Variational Auto‑Encoder (VAE) or an Implicit Neural Representation (INR) for each block. This hybrid design exploits the high‑fidelity reconstruction of INR when it outperforms the VAE.

Asymmetric Encoder‑Decoder Design

The encoder uses a deep, high‑capacity network to extract rich latent features, while the decoder employs a lightweight MobileNet‑like backbone with ShuffleNet‑style channel grouping and re‑parameterized 3×3 convolutions to keep decoding cost low. The activation function is replaced by WSiLU, a smooth, everywhere‑differentiable variant that avoids gradient truncation.

Hyper‑Prior and Context Model

A hyper‑prior network extracts side information from the latent features. A dual autoregressive context model predicts each latent element from already decoded spatial neighbors and from a non‑uniform channel partition, progressively fusing context to improve probability estimates for arithmetic coding.

Generalized Gaussian Entropy Model

The latent distribution is modeled by a generalized Gaussian with mean, scale, and shape parameters, providing a richer fit to diverse data and reducing symbol‑coding overhead.

INR Over‑fit Coding

Each image is split into eight blocks. An INR is over‑fitted to each block; if its reconstruction error is lower than the VAE’s, the learned latent variables and network weights are packed into the bitstream using a lottery‑ticket mask, enabling low‑complexity decoding while improving compression.

Subjective Quality Optimization Pipeline

Train a single‑rate model with L2 loss.

Fine‑tune for multiple rates using randomly sampled Lagrange multipliers as conditioning signals.

Add LPIPS to the loss to preserve semantic details.

Freeze encoder and entropy model, then apply a subjective aggregation loss in adversarial training to generate realistic textures.

AI‑Driven TNC Video Codec

The video codec builds on the image stack and adds:

ESRGAN‑based AI pre‑processing for detail enhancement.

Neural Network Loop Filtering (NNLF) for post‑processing.

A complexity‑aware CUTree adaptive QP calculation.

Scene detection with local complexity analysis to refine QP.

Frame‑Level Parallelism

NNLF is split into two stages: the first pass determines frame‑level parameters from a subset of rows; the second pass processes the remaining rows using those parameters, allowing dependent frames to start decoding earlier.

Algorithm‑Engine Joint Acceleration

Compute‑intensive kernels are accelerated with AVX2 intrinsics and GPU kernels. The SADL inference library gains ~45 % speedup, and adaptive block skipping based on texture and QP adds another ~45 % improvement, yielding an overall decoding speed 2.11× faster than the unoptimized baseline.

Results Summary

Image codec: average PSNR gain +1.66 dB over BPG at 0.51 bpp, peak gain +2.81 dB.

Video codec (subjective): MOS +1.20 over VTM, +0.57 over ECM.

Video codec (objective): PSNR gain +3.07 dB over VTM, +2.00 dB over ECM.

Future Outlook

While TNC demonstrates state‑of‑the‑art compression efficiency and visual quality, challenges remain for commercial deployment, including cross‑platform decoding consistency and strict terminal‑side complexity constraints. Ongoing work will focus on further algorithmic innovation and engineering optimization to bridge the gap to industry adoption.