DSAQuant Denoising-Stage-Aligned Quantization-Aware Training for Video Generation

Preserve the structure. Recover the detail. Train quantized video diffusion models in alignment with the distinct role of each denoising stage.

Paper Code Model

Simplicity

Minimal changes. Integrates quickly with existing training and inference pipelines.

Generality

Across models and scales. Works across Wan and CogVideoX at multiple sizes and resolutions.

Effectiveness

Stable gains. Averaging +2.22 VBench at INT4 and +3.34 at INT3.

02 / Motivation

Quantization errors are stage-dependent

Conventional QAT does not fail uniformly: early-stage perturbations are relatively mild, while late errors destroy detail and texture.

Observation 01 · Conventional QAT

Challenges of conventional QAT

Full-precision video frames used as the visual reference — Full-precision reference.

Uniform QAT retains the scene composition and temporal motion, but fails to recover high-frequency appearance and texture.

PreservedStructure & motion DegradedAppearance & texture

Conventional QAT results with preserved structure but degraded appearance and texture — Failure mode of conventional quantization-aware training.

Observation 02 · Stage-wise analysis

Where quantization happens matters.

The same perturbation has sharply different effects along the denoising trajectory.

Early steps Limited impact

Slight layout changes, while overall visual quality remains high.

Late steps Severe degradation

Visible artifacts emerge and fine detail breaks down.

Stage-wise comparison showing mild impact from early-step quantization and severe degradation from late-step quantization — Quantization errors affect different denoising stages differently.

03 / Method

One principle, applied twice

Quantized video diffusion models should be trained according to the role of each denoising stage.

Training

Denoising-Stage Oriented Supervision

Teacher anchoring stabilizes early structure, then smoothly yields to the native target loss so late stages can reconstruct detail.

Early stepsDenoisingLate steps

Teacher anchoring

Target learning

Inference

Denoising-Stage Gated Guidance

Classifier-free guidance is gated near the end of denoising, preventing conditional mismatch from amplifying quantization noise.

Early stepsDenoisingLate steps

CFG Active CFG Drop

Overview of the DSAQuant training and inference framework — Stage-oriented supervision during training and gated guidance during inference.

01 / Results

Low precision, high quality.

Explore eighteen INT4 DSAQuant outputs across three Wan model scales. Each collection cycles automatically between two groups of three samples.

Wan 1.3B

6 samples · INT4 DSAQuant

Seine in Van Gogh style

Panda by the campfire

Stormtrooper on the beach

Vase, zoom out

Teddy bear swimming

Jellyfish ocean

Wan 2.2 5B

6 samples · INT4 DSAQuant

Christmas fireworks

Fantasy rabbit

Panda on a swing

Raccoon guitarist

Fantasy flythrough

Bicycle ride

Wan 2.1 14B

6 samples · INT4 DSAQuant

Mountain road

Dog jumps onto couch

Garden, zoom out

Pouring tea

Handshake

Morning makeup

04 / Comparisons

Inspect every comparison

Switch between system-level comparisons and ablations, then choose a model family.

Acknowledgements

We would like to thank Yichong Lu, Jiahao Wang, Yufeng Yuan, Nan Zhou, Jiahao Shao, and Yudong Jin for their insightful discussions.