tcnerv:dual-domain temporal context modeling for implicit neural video compression
Xuezhi Xiang, Yixin Zhao, Heqi Xiang, Jiayao Liu, Shanjun Zhang
cs.CV
Quick verdict
TCNeRV adds dual-domain temporal context to INR video coding; +2.2 dB vs HNeRV-Boost on UVG at 3M params; credible, but single-dataset and fades at 10M scale.
Summary
TCNeRV is a per-video implicit neural representation (INR) video codec built on HNeRV-Boost that exploits temporal redundancy in two places. First, multi-scale temporal-context fusion (MTCF) resizes the previous reconstructed frame, projects it through a small convolutional block, and fuses it via learnable channel-wise sigmoid gates into decoder Stages 3 and 4 (Sec. 2.2). Second, temporal embedding-residual coding (TERC) predicts each frame's content embedding from a stop-gradient encoding of the previous reconstruction with a 1x1-conv predictor, quantizing and entropy-coding only the residual; the first frame is coded in full (Sec. 2.3). On UVG at ~3M parameters, TCNeRV reaches 36.08 dB average PSNR, +2.20 dB over HNeRV-Boost and +0.82 dB over HiNeRV; gains versus HNeRV-Boost persist at 6M (+1.70 dB) and 10M (+1.11 dB) scales, though HiNeRV surpasses TCNeRV at 10M (38.02 vs 37.60 dB). Reported BD-rate reductions are 57.26% vs HNeRV-Boost, 29.85% vs HiNeRV, 22.06% vs HM, and 66.73% vs DCVC, but TCNeRV is worse than DCVC-DC (+8.31%) and NVRC (+40.22%). Ablations attribute most of the gain to MTCF (+2.08 dB). Main limitations: evaluation confined to UVG's seven sequences, no FFNeRV/DNeRV comparison, regression-only ablation, unquantified autoregressive error accumulation, and no code or runtime analysis.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Injecting gated previous-reconstruction features at two decoder scales (MTCF) plus embedding-residual coding (TERC) yields +2.20 dB average PSNR over HNeRV-Boost on UVG at ~3M parameters.
- TCNeRV reduces BD-rate by 22.06%, 66.73%, and 29.85% relative to HM, DCVC, and HiNeRV respectively on UVG.
- MTCF provides the primary reconstruction gain (33.88 to 35.96 dB) while TERC adds a further 0.12 dB, at a cost of only 0.01M parameters.
- Exploiting temporal context in both the decoding-feature domain and the content-embedding domain outperforms exploiting either alone.
Strengths
Well-motivated dual-domain decomposition with negligible overhead
MTCF and TERC target two distinct limitations stated in Sec. 1 (unconditioned decoding features; independently coded embeddings) and together add only 0.01M parameters and 0.4G MACs over the baseline (Table 3: 3.05M/130.6G to 3.06M/131.0G).
Large, consistent regression gains at matched model budget
Table 1: +2.20 dB over HNeRV-Boost at 3M (36.08 vs 33.88 dB), with the largest per-sequence gains exactly where temporal context should help high-motion content (Ready +5.45 dB, Yacht +3.79 dB); gains of +1.70/+1.11 dB also hold at 6M/10M scales.
Honest reporting of unfavorable comparisons
Clean component attribution via ablation
Weaknesses
Single-dataset evaluation
Scale-dependent advantage over HiNeRV is not disclosed in claims
Ablation never measures the compression metric
Closest temporal-INR baselines omitted
Cost of the causal loop unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TCNeRV?
TCNeRV is an implicit neural representation (INR) video compression method that builds on HNeRV-Boost and reuses the previously reconstructed frame in two ways: gated multi-scale feature fusion into the decoder (MTCF) and prediction-residual coding of per-frame content embeddings (TERC), so only temporally unpredictable information is stored.
How much better is TCNeRV than HNeRV-Boost?
On the UVG dataset with about 3M parameters, TCNeRV averages 36.08 dB PSNR versus 33.88 dB for HNeRV-Boost (+2.20 dB), and in BD-rate terms its compressed bitstream is 57.26% smaller at equal quality.
Does TCNeRV beat standard codecs like HEVC or DCVC?
The paper reports lower BD-rate than HM (22.06% less bitrate) and the original DCVC (66.73% less), but TCNeRV is still 8.31% worse than the newer DCVC-DC and 40.22% worse than NVRC, so the latest neural video codecs remain ahead.
Is TCNeRV code publicly available?
The paper does not mention a code release. Reproduction relies on the described architecture and settings: the HNeRV-Boost pipeline with a ConvNeXt encoder, Adan optimizer, 300 training epochs plus 100 compression fine-tuning epochs, and 8-bit quantization with a Gaussian entropy model.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are present only as garbled extracted text; the RD curves in Fig. 4 could not be visually inspected, so overlap/consistency with Table 2 could not be checked.
- Reference [21] (NIRVANA) shows a malformed author list ('S. R. et al'), which may be an extraction artifact rather than an error in the paper.
- All quantitative judgments rely on the numbers as extracted from Tables 1-3; no independent recomputation of BD-rate or PSNR values was possible.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.