Temporal-Consistent Quantization for Video Stability
Implement temporal-aware INT8 quantization across 3D blocks to eliminate precision errors and flickering artifacts in video generation.
Time
Compute level
Resources
Prerequisites
Scope
The Gap
Section 1.3 describes 'block-wise granularity of 128 x 128' for INT8 quantization. This spatial-only blocking ignores the temporal dimension of video data.
Limitation
Quantizing activations independently per spatial block can cause quantization scales to fluctuate between frames, introducing 'quantization flicker' (temporal noise) that degrades video quality despite acceptable single-frame metrics.
Proposed Solution
Implement Temporal-Aware Quantization (TAQ). Modify the quantization kernels to calculate scales across 3D blocks (Time x Height x Width) or smooth scaling factors temporally to ensure consistent precision errors across frames.
Why This Matters
Ensures that the 200x speedup does not come at the cost of temporal stability, which is critical for professional video generation workflows.
How to Get Started
- Analyze quantization error propagation across the temporal dimension in existing INT8 models.
- Modify the quantization kernel to compute scaling factors over 3D spatio-temporal blocks.
- Apply temporal smoothing constraints to quantization parameters during calibration.
- Evaluate video stability and visual quality metrics against spatial-only quantization baselines.
Code linked to this paper on Hugging Face: github.com/thu-ml/TurboDiffusion
Suggested by AI. Adapt it to your setup.
Has This Been Done?
Papers such as Quantizing Diffusion Models and TinyNeRV investigate low-precision inference and quantization for generative models, but none propose Temporal-Aware Quantization (TAQ) using 3D block-based scales or temporal smoothing for video stability.
- Quantizing Diffusion Models for Scalable and Efficient Generative Inference Across Diverse Hardware Platforms · 2025 · 0 citations
- TinyNeRV: Compact Neural Video Representations via Capacity Scaling, Distillation, and Low-Precision Inference · 2026 · 0 citations
We searched OpenAlex for “Implement temporal-aware INT8 quantization across 3D blocks to eliminate precision errors and flickering artifacts in video generation.” and “Temporal-Consistent Quantization for Video Stability” and an AI compared the idea with the top 10 results. A small search can miss relevant work.
Source Paper
TurboDiffusion: Accelerating Video Diffusion Models by 100-200 TimesJintao Zhang, Kaiwen Zheng, Kai Jiang +5 more · 2025
The paper introduces TurboDiffusion, a framework designed to accelerate Wan video diffusion models by 100-200x on NVIDIA RTX 5090 GPUs. It achieves this by combining four optimizations: low-bit attention (SageAttention), sparse-linear attention (SLA), step distillation via regularized Consistency Models (rCM), and W8A8 quantization of linear layers.
Related Ideas
- Temporal-Consistency Enhanced Super-ResolutionHunyuanVideo 1.5 Technical Report🟠Medium
- Spatiotemporal SpotEdit for Video DiTsSpotEdit: Selective Region Editing in Diffusion Transformers🔴Hard
- Dynamic Sub-Second Boundary Alignment for Asymmetric ModalitiesOmniForcing: Unleashing Real-time Joint Audio-Visual Generation🔴Hard
- Spatiotemporal Prism: Unified Video AutoencodingThe Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding🔴Hard
- Motion-Aware Transition Modeling via Multi-Frame ConditioningStoryMem: Multi-shot Long Video Storytelling with Memory🟡Fairly easy