- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
PSC aims to accelerate autoregressive image and video generation without an auxiliary draft model or a change to the target sampling distribution. Its starting point is that proposal distributions recycled from earlier speculative rounds lag behind the current context, reducing proposal–target overlap. PSC extrapolates changes in aligned distributions to predict a better proposal. A first maximal coupling transports an existing sample to this proposal, and a second corrects it to the target distribution. The method combines this mechanism with vectorized decoding and adaptive prediction strength. The paper presents a distribution-preservation argument and evaluates Lumina-mGPT-7B, Janus-Pro-7B, and Cosmos-AR-4B, reporting fewer target-model evaluations and lower latency than SJD while maintaining comparable measured generation quality.
Critical strengths
Vectorization yields substantial acceleration. Table 4 reduces latency from 13.8091 to 6.8262 seconds with little NFE change, and the no-prediction control helps identify its source. This is the clearest demonstrated practical value.
Critical weaknesses
-
Prediction has limited demonstrated value beyond vectorization. On Janus, NFE is almost unchanged and prediction is faster at window 64 but slightly slower at window 32. On Cosmos, PSC is 9%–15% slower than MC at all four windows; Lumina lacks a matched vectorized no-prediction control. Moreover, at Janus window 32, MC NFE changes from 241.4704 in Table 2 to 153.3814 for Scalar MC in Table 6, while PSC NFE remains 152.7910; PSC timings also differ. These issues make it difficult to attribute the reported acceleration to the central predictive mechanism.
-
The distribution-preservation argument leaves a conditional step unexplained. Composing the two maximal couplings is valid under the required assumptions. The unresolved point is why a reused token still follows q conditional on adaptive predictor history and selection. Appendix A.2 asserts this conditional validity without establishing it. Because exact preservation is a central claim, this step matters to the overall assessment; it is a proof-completeness concern, not evidence of demonstrated sampling bias.
Questions for authors
-
When does prediction outperform matched β=0 decoding? Please extend the control to Lumina and Cosmos and report tokens committed per target call, total tokens processed, and the frequency and benefit of nonzero β choices.
-
Please reconcile Tables 2, 4, and 6, including differences in samples or timing conditions. Repeated timings and a breakdown of target-forward, prediction, and coupling costs would clarify whether prediction offsets its overhead.
-
Why does the reused token retain the conditional law required by Theorem 2 under adaptive history and selection? Please specify the dependencies and show how the condition is maintained across rounds.