Source-linked AI summary
Information Capacity of Generative Video Compression: Quantifying the Rate-Compute Exchange at Identical Quality
Cheng Yuan, Jiawei Shao, Xuelong Li
TL;DR
The rate–compute exchange at identical reconstruction quality has not been quantified, limiting comparisons of decoder compute and bandwidth savings. The paper fits a two-factor quality surface and defines information capacity along iso-quality contours; across five datasets, the 14B decoder converts compute into rate savings roughly ten times more efficiently than the 1.3B decoder.
Problem
The exchange between data rate and decoder complexity at constant output quality is not captured by standard codec comparisons, leaving the rate savings from additional decoder compute unquantified.
Source: §1 · p. 1Method
The paper fits a two-factor power-law quality surface in rate and decoder compute, then defines information capacity as the negative logarithmic slope along an iso-quality contour.
Source: §1 · p. 1, §2.2 · p. 2Results
Across five datasets, the 14B decoder converts compute into rate savings roughly ten times more efficiently than the 1.3B decoder, and information capacity varies significantly across datasets.
Source: §1 · p. 1Takeaways & Limitations
Because information capacity varies with rate and compute and is unit-invariant, it can identify operating points where extra compute offers a favorable rate exchange and support decoder comparisons.
Source: §2.2 · p. 2Compute is measured as theoretical DiT FLOPs rather than deployment latency or energy, and the reported information capacity uses DISTS alone.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Under the AI Flow framework, communication networks distribute intelligence across devices, edge servers, and clouds, and computation at the receiver becomes a resource that can substitute for transmitted bits. Generative video compression (GVC) embodies this exchange by sending compact tokens with ultra-low bitrate and letting a generative decoder synthesize the video, yet how much bandwidth savings a unit of decoder compute actually achieves has never been quantified. To fill this vacancy, we model reconstruction quality as a two-factor power law in data rate and decoder compute, which fits measured DISTS of two GVC decoders with a mean error below 3%, and define the information capacity (IC) as the negative logarithmic slope along an iso-quality contour, namely the fraction of rate saved per fractional increase in compute at identical quality. IC is dimensionless and unit-invariant, thus enabling an architecture-agnostic comparison. It forms a field over the operating plane, locating where additional denoising steps are worth their cost. Across five datasets, the 14B decoder trades more compute for fewer rate about ten times more efficiently than the 1.3B decoder. IC also varies significantly across datasets, indicating imbalanced performance on the rate-compute trade-off in GVC methods.
1. INTRODUCTION
GVC makes decoder computation a potential substitute for transmitted bits, but the rate–compute exchange at equal quality has not been quantified. This paper defines information capacity for that exchange and reports that the 14B decoder is roughly ten times more efficient than the 1.3B decoder at converting compute into rate savings.
1. INTRODUCTION
- The rate–compute exchange at constant quality has not been quantified, leaving questions such as decoder compute allocation under an uplink budget unanswered.Standard BD-rate comparisons measure rate savings at equal quality but ignore decoder complexity.Source: §1 · p. 1
- The paper models quality as a two-input function of rate and decoder compute and defines information capacity as the negative logarithmic slope along an iso-quality contour.This makes the rate savings per fractional increase in compute measurable at identical quality.Source: §1 · p. 1
- The 14B decoder converts compute into rate savings roughly ten times more efficiently than the 1.3B decoder, while information capacity varies significantly across datasets.The study fits measured DISTS for two decoders across five datasets with mean error below 3%.Source: §1 · p. 1
2. METHODS
The methods represent GVC quality as a function of bitrate and decoder compute, then quantify the rate reduction available from additional compute along equal-quality contours. The resulting information capacity is unit-invariant and varies across the operating plane.
2.1. Rate–compute–quality surface
- A GVC configuration is characterized by transmitted bitrate, decoder compute, and reconstruction distortion, with rate controlled by keyframe QP and compute by DiT denoising steps in this study.The formulation is intended to apply to GVC implementations exposing configurable rate- and compute-related inference hyperparameters.Source: §2.1 · pp. 1–2
- The two-factor power law models distortion as separate rate- and compute-dependent terms plus a floor, assuming the resources remove largely distinct distortion components with diminishing returns.The rate term captures distortion removed by more bits, the compute term by additional decoder compute, and the floor represents irreducible distortion.Source: §2.1 · p. 2
2.2. Information capacity
- Information capacity is the negative logarithmic slope of an iso-quality contour, measuring how rate changes with compute while quality remains fixed.The paper uses the closed form derived by implicit differentiation of the fitted surface, which smooths measurement noise and extends the contour across the measured rectangle.Source: §2.2 · p. 2
- A one-percent compute increase permits an information-capacity-percent reduction in rate at unchanged quality; for a compute multiplier k, η_k gives the saved rate fraction.For k = 2, η_2 is the rate saved by doubling the decoder budget.Source: §2.2 · p. 2
- Information capacity is dimensionless and unit-invariant, and its value depends on both rate and compute, so any scalar comparison must specify an operating point.It is largest at high rate and low compute; the figure presents the field across datasets and decoder sizes.Source: §2.2 · p. 2, Figure 2 · p. 3
3.1. Experimental setup
The experiments compare two GVC decoders across controlled rate and compute settings, fitting a DISTS quality surface for each model–dataset pair. The setup uses five datasets and reports fitted parameters and information capacity alongside fit error.
3.1.1. Models
- The evaluation uses GVC-1.3B and GVC-14B, which reconstruct GOPs with Wan2.1-based DiT decoders conditioned on decoded keyframes.The 1.3B system additionally sends a sparse canny edge map; the decoders differ in parameter count and GOP length.Source: §3.1.1 · p. 3
3.1.2. Empirical measurements
- Keyframe QP controls rate, while denoising-step count controls compute; the measured DiT cost excludes small auxiliary components independent of step count.The measurements use 88 configurations per dataset for GVC-1.3B and 64 for GVC-14B.Source: §3.1.2 · p. 3
3.1.3. Fitting procedure
- The fit estimates five surface parameters by relative-error-weighted nonlinear least squares, using geometric-center rate and compute normalizers.The normalizers make the fitted rate- and compute-limited contributions directly interpretable at the center of the measured grid.Source: §3.1.3 · p. 3
3.1.4. Metric and dataset
- Quality is measured with DISTS, and one surface is fitted per model and dataset using mean scores over five datasets.The datasets comprise HOIGen-1M, MCL-JCV, SA-V, TemporalBench, and Sintel.Source: §3.1.4 · p. 3
3.2. Fitted surfaces
The two-factor power law fits measured DISTS closely across the evaluated configurations. Its fitted exponents distinguish the decoders’ sensitivity to compute, while rate exponents and distortion floors vary across dataset–model pairs.
3.2. Fitted surfaces
- Mean relative fit error is 0.4%–2.7%, with maximum error below 7.1% across 760 configurations.Table 1 lists fitted parameters and errors; Fig. 1 shows the fitted DISTS surface on MCL-JCV.Source: §3.2 · p. 3
- For nine of ten pairs, rate exponent b ranges from 1.2 to 3.6; it reaches 6.8 for GVC-1.3B on Sintel, whose rate range is narrowest, a factor of 1.24.Source: §3.2 · p. 4
- The fitted floor e is 0.05–0.12 for eight pairs and vanishes for GVC-1.3B on HOIGen-1M and TemporalBench.For those two pairs, the rate term alone accounts for residual distortion within the measured range.Source: §3.2 · p. 4
- The compute exponent d ranges from 0.06 to 0.73 for GVC-1.3B and 0.57 to 2.02 for GVC-14B.The larger decoder is strongly compute-limited at few steps, with contours bending toward the rate axis only once N exceeds roughly 4.Source: §3.2 · p. 4
3.3. Information capacity over the operating plane
Information capacity varies across both operating points and datasets, identifying where extra compute can save rate. The favorable exchange regime is much more broadly reached by GVC-14B than by GVC-1.3B in the measured range.
3.3. Information capacity over the operating plane
- Within each dataset–model pair, IC varies by a factor of five to several hundred across bitrates and FLOPs, and is largest at high rate and low compute.The field therefore distinguishes favorable operating regions that a single scalar cannot identify.Source: §3.3 · p. 4
- GVC-14B reaches IC ≥ 1 within the measured grid on four datasets and exceeds it everywhere on SA-V; GVC-1.3B reaches it only on SA-V.For GVC-1.3B, that regime occurs only at the highest rates and fewest steps; on the other four datasets IC stays below 0.5 throughout.Source: §3.3 · p. 4
3.4. Cross-model and cross-dataset comparisons
Across five datasets, the 14B decoder has substantially higher information capacity than the 1.3B decoder, although capacity varies with dataset and operating point. Dataset differences align with how effectively additional bitrate reduces distortion.
3.4. Cross-model and cross-dataset comparisons
- IC0 is evaluated at the geometric centers of measured rate and compute ranges, representing a typical operating point rather than a grid extreme.The paper reports IC0 together with its range over the measured grid.Source: §3.4 · p. 4
- IC0 is 6–33 times higher for GVC-14B than GVC-1.3B across datasets, with a median factor of 9 and ranges of 0.46–4.16 versus 0.05–0.66.The 14B decoder’s IC0 is more sensitive to operating point, but it consistently exceeds the 1.3B decoder under identical parameter configurations.Source: §3.4 · p. 4
- Doubling decoder compute at each model’s grid center saves 3.5–36.8% of rate for GVC-1.3B and 27.4–94.4% for GVC-14B.The two decoders were measured over different rate ranges, and each IC0 refers to its model’s own operating center on a dataset.Source: §3.4 · p. 4
- SA-V has the highest IC0 and HOIGen-1M the lowest for both decoders; their rate exponents are approximately 1.3 and at least 2.5, respectively.The paper attributes this ranking largely to the rate exponent: bitrate is less effective at removing distortion on SA-V and more effective on HOIGen-1M.Source: §3.4 · p. 4
4. CONCLUSION
The paper defines information capacity as rate elasticity with respect to decoder compute at constant quality and applies it to two GVC decoders across five datasets. The 14B decoder is about ten times more efficient at converting compute into rate savings, while conclusions are bounded by the compute measure and quality metric used.
4. CONCLUSION
- Across two GVC decoders and five datasets, the 14B decoder converts compute into rate savings about ten times more efficiently than the 1.3B decoder.The fitted surface matches measured DISTS to within a few percent errors, and the exchange rate varies across operating points and datasets.Source: §4 · p. 4
- Information capacity is the elasticity of rate with respect to decoder compute along an iso-quality contour, derived in closed form from a two-factor power-law surface.Source: §4 · p. 4
- The conclusions are bounded by theoretical DiT FLOPs, which omit memory-bound execution, latency, and energy, and by reporting IC only for DISTS.The same IC may correspond to different deployment costs on heterogeneous hardware; other quality metrics may yield different fitted exponents.Source: §4 · p. 4
- Future work will test more GVC decoders and architectures, assess downstream-task performance, and investigate IC for adaptive step allocation across device, edge, and cloud nodes.Source: §4 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The study is a numerical simulation and did not require ethical approval.
6. COMPLIANCE WITH ETHICAL STANDARDS
- No ethical approval was required because this was a numerical simulation study.Source: §6 · p. 5