600倍加速，720p视频实时生成！单卡也能带的动14B模型
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - LightX2V released LightWan2.2-A14B, an inference-optimization stack for the 14B Wan2.2-A14B video diffusion model that claims up to ~705x speedup, bringing 5-second 720p generation down to 3.8s (T2V) / 4.5s (I2V) on 8x RTX 5090 and making the model runnable on a single consumer GPU. It matters because it pushes high-resolution video generation from tens of minutes into real-time, reproducible territory on consumer hardware.
- Step reduction: Phased DMD splits the SNR range across stages (two high-noise + two low-noise expert steps) to cut denoising from 40 steps to 4, with SGMD aligning fake score to teacher score for faster, more stable distillation.
- Per-step compression: NVFP4 quantization-aware training (E2M1 4-bit, FP8 E4M3 block scales per 16 elements) folded into distillation, plus dynamic sparse attention keeping only the top ~10–20% of Q/K blocks — self-attention otherwise dominates >80% of DiT latency at ~120K tokens.
- Systems work: CUTLASS block-scaled NVFP4 GEMMs on Blackwell, fused 3D RoPE/RMSNorm/Triton kernels, Dynamic Sparse SageAttention (Q/K INT8, V FP8), and block-level asynchronous offload overlapping weight transfer with compute to fit 14B weights under ~30GB.
- Reported numbers: single-card T2V 720p 2668s → 22.5s (118.7x), I2V 720p 2685s → 26.7s (100.5x), ~3.5x/2.4x faster than TurboWan2.2 at 480p/720p; multi-GPU uses Light-Ulysses sequence parallelism with FP8 All-to-All, QKV tensor fusion, and head-level pipelining.
Sources (1)
600倍加速，720p视频实时生成！单卡也能带的动14B模型
TL;DR - LightX2V released LightWan2.2-A14B, an inference-optimization stack for the 14B Wan2.2-A14B video diffusion model that claims up to ~705x speedup, bringing 5-second 720p generation down to 3.8s (T2V) / 4.5s (I2V) on 8x RTX 5090 and making the model runnable on a single consumer GPU. It matters because it pushes high-resolution video generation from tens of minutes into real-time, reproducible territory on consumer hardware.
- Step reduction: Phased DMD splits the SNR range across stages (two high-noise + two low-noise expert steps) to cut denoising from 40 steps to 4, with SGMD aligning fake score to teacher score for faster, more stable distillation.
- Per-step compression: NVFP4 quantization-aware training (E2M1 4-bit, FP8 E4M3 block scales per 16 elements) folded into distillation, plus dynamic sparse attention keeping only the top ~10–20% of Q/K blocks — self-attention otherwise dominates >80% of DiT latency at ~120K tokens.
- Systems work: CUTLASS block-scaled NVFP4 GEMMs on Blackwell, fused 3D RoPE/RMSNorm/Triton kernels, Dynamic Sparse SageAttention (Q/K INT8, V FP8), and block-level asynchronous offload overlapping weight transfer with compute to fit 14B weights under ~30GB.
- Reported numbers: single-card T2V 720p 2668s → 22.5s (118.7x), I2V 720p 2685s → 26.7s (100.5x), ~3.5x/2.4x faster than TurboWan2.2 at 480p/720p; multi-GPU uses Light-Ulysses sequence parallelism with FP8 All-to-All, QKV tensor fusion, and head-level pipelining.