Video compression at ultra-low bitrates remains challenging due to the loss of perceptually important details under limited bit budgets. Recent generative video compression methods leverage pretrained video priors to improve perceptual quality and temporal coherence, but many follow a compress-then-generate paradigm, where generative priors are primarily exploited during reconstruction. Moreover, representations learned for generation and reconstruction are not inherently optimized for entropy coding. In this work, we propose \method, which adapts pretrained video representations into compression-oriented representations while preserving their reconstruction capability. Specifically, we adapt the pretrained causal video VAE of MiniMax-H3 into a discrete video tokenizer and develop a small autoregressive entropy model that captures local and global token dependencies, with grouped prediction for parallel decoding. To further improve coding efficiency, we jointly optimize the tokenizer and entropy model using a differentiable rate surrogate, allowing rate gradients to adapt token assignments toward lower coding costs while preserving reconstruction quality. Experiments on UVG, MCL-JCV, and HEVC Class B demonstrate superior perceptual rate--distortion performance, achieving BD-LPIPS savings of 82.2%--87.5% and BD-DISTS savings of 78.3%--85.0% relative to GLC-Video.
Figures & tables
Figure 1: h3vc achieves average BD-rate savings of 84.1% and 82.3% on UVG, MCL-JCV, and HEVC Class B over GLC-Video [ 21 ] , at same LPIPS and DISTS.
Figure 2: Overview of h3vc . Our H3 encoder transforms the input video x into continuous latents y , which are quantized into discrete indices z using vector quantization. These indices will be entropy coded and decoded through our categorical entropy model. At the decoding side, the recovered indices z are mapped to quantized latents y^ through codebook lookup, which are then passed to our H3 decoder to reconstruct x^ . During Stage III, the encoder, decoder, and entropy model are jointly optimized with the codebook fixed; rate gradients reach the encoder through soft assignments.
Figure 3
Coding
bits/token
w/o Entropy Coding
12.000
w/ Entropy Coding
8.487
Table 2: Coding cost before and after stage II entropy model training.
Figure 4: Window-parallel autoregressive entropy modeling. Token groups follow the frame-major order r=4t+g , with spatial windows processed in parallel. Group-causal attention captures local dependencies, while cross-attention incorporates global context retrieved from a memory bank of previously decoded frames.
Figure 5: Token predictability under different rate constraints on UVG. (a) Codebook usage ( ∙ ) and top-1 prediction accuracy ( ▲ ). (b) Coding cost (bits/token): the bit cost assigned by the entropy model to each observed token.
Figure 6: Perceptual rate–distortion performance against state-of-the-art video codecs on HEVC-B, UVG, and MCL-JCV.
Figure 7: Qualitative comparison on MCL-JCV. Best viewed when zoomed in.
Figure 8: Effect of rate gradients on joint rate–distortion training. Percentages show the evaluation-loss reductions at 20K steps.
Context
HEVC-B
UVG
MCL-JCV
LPIPS
DISTS
LPIPS
DISTS
LPIPS
DISTS
w/o local context
+6.83
+15.95
+1.99
+19.01
+5.69
+13.01
w/o global context
+5.47
+12.36
+0.64
+17.76
+4.42
+11.02
Table 4: Local and global context ablation. BD-rate (%) relative to H3VC at matched LPIPS-VGG and DISTS.
Figure 9: Test-time rate adjustment via prefix transmission. LPIPS, DISTS, and PSNR curves on UVG, with Bosphorus reconstructions (bpp / LPIPS). Full H3VC and H3VC-Prefix use the same checkpoint.
Temporal quality
Method
Rate (bpp) ↓
Ewarp↓
CLIP-F ↑
HEVC
0.00636
27.29
0.9907
VVC
0.01017
25.43
0.9921
DCVC-RT
0.00626
24.74
0.9915
GLC-Video
0.00775
128.98
0.9884
GVC1D
0.00821
65.26
0.9908
Table 6: Temporal reconstruction quality on HEVC-B. Bold indicates the best value in each column.