acceptodds
Under review as a conference paper at ICLR 2027

TrajeQ: Trajectory-Preserving Flow Matching Quantization for Low-Bit Video Generation

Abstract

While flow matching (FM) has emerged as a prominent paradigm for high-quality video generation, its significant computational overhead presents a major obstacle to practical implementation. Although post-training quantization (PTQ) is an established approach for efficient generation, existing PTQ methods primarily minimize reconstruction error in a magnitude-centric manner. This objective is misaligned with deterministic FM trajectories, where low-bit quantization introduces velocity-field orientation errors that can compound across integration steps and cause deviations from the full-precision trajectory. To address this issue, we introduce ***TrajeQ*** to preserve the geometric integrity of the velocity field and reduce trajectory drift under low-bit quantization. Our approach introduces *Curvature-Adaptive Quantization (CAQ)* to detect angular shifts across timesteps and mitigate geometric distortion via adaptive precision. We also integrate *Directional Bottleneck Preservation (DBP)*, a sensitivity-guided precision assignment scheme that retains only the most vulnerable layers in full precision. Experiments demonstrate that ***TrajeQ*** achieves state-of-the-art VBench performance and maintains stable video generation under extreme W4A4 and W3A6 quantization regimes where existing methods degrade severely. We provide the code in the supplementary material and additional video results at https://anonymous-orange.github.io/.

Primary area:
generative models
Keywords:
Video Generation, Flow-Matching, Quantization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dQeJOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem that, in low-bit post-training quantization of flow-matching video models, quantization errors change the velocity direction and accumulate along the denoising trajectory, causing generation to collapse. It proposes to temporarily raise the activation bit-width at timesteps with large direction changes (CAQ), and to keep a few key layers in FP16 according to their directional sensitivity (DBP). On Pyramid Flow, it shows better generation quality than MinMax, SmoothQuant, QuaRot, Q-DiT and ViDiT-Q from W6A6 to W3A6. However, it does not verify that the errors at the selected locations actually have the largest impact on the final result, it lacks comparisons with SVDQuant/DeltaQuant and real-deployment efficiency results, and the evaluation is small, with each configuration run only once.

Critical strengths

1、The code is anonymously open-sourced.

2、 The motivation deserves recognition, and it is supported by experiments.

Critical weaknesses

1、The paper is motivated by “errors propagating and accumulating along the denoising trajectory”, but both CAQ and DBP allocate precision only according to local direction changes, without verifying whether the errors at these locations actually have the largest impact on the final video.

2、Several descriptions in the paper need further clarification:

(1) The profiling set used by DBP for layer selection is not specified in terms of size, source and seeds, which makes it hard to rule out systematic bias.

(2) How exactly is the CAQ threshold determined? The paper describes using 10 prompts, but Tables also report results on these same 10 prompts. Is this evaluation reasonable?

(3) The main text states that quantization scales are computed on the fly without offline calibration, whereas Appendix A.1 describes a protocol based on a calibration set. Please clarify the details.

3、The paper claims SOTA under extreme settings such as W4A4, yet excludes SVDQuant and DeltaQuant, which also target 4-bit diffusion/video quantization, on the grounds that they are “a different type of method”. Given that TrajeQ itself also keeps some layers in FP16, this reason for exclusion does not hold. Comparisons should be added under the same bit-width or the same memory/compute budget.

Questions for authors

pls see weeknesses.

Reviewer_iwDFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a new post-training quantization method for flow matching video generation models called TrajeQ. The main idea is that small quantization errors can accumulate during sampling by changing the direction of the predicted velocity field, leading to trajectory drift. To alleviate this, the method preserves direction sensitive transformer layers via Directional Bottleneck Preservation (DBP) and allocates higher precision to important sampling timesteps identified through trajectory curvature termed Curvature Adaptive Quantization (CAQ). Experiments on several video generation benchmarks show consistent improvements over existing PTQ work under low-bit quantization settings such as W4A4 and W3A6.

Critical strengths
  1. The main strength of the paper is its trajectory-preservation perspective on quantizing flow matching models. The observation that directional errors in the predicted velocity can accumulate during sampling provides a clear motivation for the method. Based on this insight, the proposed DBP and CAQ components allocate precision across layers and timesteps, respectively, and yield consistent improvements over existing PTQ baselines across several benchmarks and low-bit quantization settings.
Critical weaknesses
  1. The evaluation relies entirely on automated metrics and qualitative examples, without any human preference study. Since the paper makes claims about preserving perceptual video quality under aggressive quantization, a user study would help verify that the reported metric improvements translate into noticeable improvements for human viewers.
  2. The relationship between TrajeQ and existing PTQ techniques could be clarified further. Since CAQ and DBP focus on allocating precision across timesteps and layers, they appear potentially complementary to methods that address activation outliers, scaling, or quantization granularity. However, the paper does not evaluate whether combining these approaches provides additional gains.
Questions for authors
  1. Can the authors provide a human preference study comparing TrajeQ with the other relevant PTQ baselines?

  2. Can CAQ and DBP be combined with existing PTQ methods such as SmoothQuant, QuaRot, Q-DiT, or ViDiT-Q, and do these combinations provide further improvements?

  3. Does TrajeQ remain effective in more extreme 1-2 step generation settings with step-distilled models?

Reviewer_YuWeOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TrajeQ addresses post-training quantization of flow-matching video generators. CAQ uses the cosine distance between activations at successive sampling steps to decide when to increase activation precision by one bit. DBP ranks layers by the ratio of angular error to squared reconstruction error and retains selected layers in FP16 under a fixed parameter budget. Experiments on Pyramid Flow miniFLUX, MMDiT, and CogVideoX report generation quality and trajectory diagnostics, with additional few-step and resolution studies. Low-bit inference is evaluated through simulated quantization using floating-point operations.

Critical strengths
  1. The low-bit quality gains are substantial and supported by direct trajectory measurements. At W3A6, Table 1 reports subject/scene consistency of 90.23/38.29, versus 81.39/31.06 for the best competing results. Table 5 also reduces angular error from 19.33 to 10.96 degrees and trajectory deviation from 0.755 to 0.452 relative to ViDiT-Q. These measurements support the trajectory-preservation claim beyond conventional temporal consistency scores.

  2. The main comparisons control precision allocation. Table 3 matches the number of precision-boosted steps, Table 10 fixes the FP16 parameter budget at 0.88%, and Section 4.1 states that the main baselines retain the same FP16 layers. These controls make the quality gains more convincing than a comparison with unmatched high-precision budgets.

Critical weaknesses
  1. The practical resource benefit remains unestablished. Section 4.1 and Table 16 use simulated quantization; the reported 0.71 GB model size is a bit-width estimate rather than measured integer-model storage. Measured memory and generation time are 42,670 MB and 35.2 h, compared with 37,499 MB and 33.5 h for ViDiT-Q, without a matched FP16 reference. CAQ also requires cached activations, and the call-weighted 6.03-bit average in Appendix E.5 does not account for varying activation sizes. The paper demonstrates better quality under simulated low precision, but not yet the end-to-end savings motivating its use on constrained devices.

  2. DBP's directional criterion is not separated from its preference for low-norm representations. In Eq. (7), scaling both reference and quantized outputs by a factor a leaves the angular error unchanged but scales DS by 1/a², ignoring the stabilizer. Appendix B.6, Eq. (37), explicitly contains an inverse-squared-norm factor. Table 10 compares MSE and angular-only rankings but not a simple inverse-norm ranking. DBP works empirically, but the evidence does not establish how much of its benefit comes from identifying directional bottlenecks rather than representation scale.

Questions for authors
  1. Could you add a matched FP16 reference to Table 16 and separate memory used by weights, current activations, CAQ caches, and temporary buffers? Please also report an element-weighted activation bit width and distinguish measured costs from estimates. This would clarify the resource trade-off without requiring new integer kernels.

  2. Under Table 10's FP16 budget, how does an inverse-squared-activation-norm ranking compare with DBP? A small matched-budget experiment would help determine whether the directional term adds value beyond the scale factor.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.