UCSD Lab Releases FastWan-QAD Video AI That Generates 5-Second Clips in 1.8 Seconds
FastWan-QAD brings near-instantaneous local video generation to consumer hardware, cutting generation time from minutes to under two seconds on a single GPU.
Reporting from 1 sources: GIGAZINE.
Hao AI Lab at UC San Diego released FastWan-QAD, a video generation AI model that produces a 480p 5-second video in 1.8 seconds on an RTX 5090. The model applies quantization-aware distillation to Wan2.1-T2V-1.3B, achieving up to 95.6x speedup over the base model while suppressing quality loss. Three variants are available under Apache License 2.0.
Hao AI Lab at UC San Diego released FastWan-QAD, a video generation AI model that runs on a single GeForce RTX 5090 and produces a 480p 5-second clip in 1.8 seconds. The model applies a method called quantization-aware distillation (QAD) to the Wan2.1-T2V-1.3B base model, reducing computational precision to speed inference while limiting quality loss. Three variants are available: FastWan-QAD-1.3B (NVFP4 precision, SageAttention3 FP4 backend, 1.8 seconds), FastWan-QAD-1.3B-SA2 (higher quality, SageAttention2++, 2.01 seconds), and FastWan-QAD-FP8-1.3B (FP8 precision, runs on RTX 40 series, 3.4 seconds on an RTX 4090). The fastest variant achieves a 95.6x speedup over the base model. All three are distributed under the Apache License 2.0.
- FastWan-QAD-1.3B: NVFP4 precision, SageAttention3 FP4 backend, 480p 5-second video in 1.8 seconds on RTX 5090, 6.4x speed over base model
- FastWan-QAD-1.3B-SA2: SageAttention2++ backend, higher quality than 1.3B, 480p 5-second video in 2.01 seconds on RTX 5090, 5.6x speed over base model
- FastWan-QAD-FP8-1.3B: FP8 precision, runs on RTX 40 series, 480p 5-second video in 3.4 seconds on RTX 4090
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.