🚀 Benchmark¶
On this page
These are historical measurements. The original records do not provide complete run dates and commit IDs. Latency and memory figures describe the listed test environments, not current-version performance; software versions identify the test conditions. Speedups are calculated from the displayed times, relative to Wan2.1 Official on H200 and Wan2GP on RTX 4090 at the same resolution.
This document showcases the performance test results of LightX2V across different hardware environments, including detailed comparison data for H200 and RTX 4090 platforms.
🖥️ H200 Environment (~140GB VRAM)¶
📋 Software Environment Configuration¶
| Component | Version |
|---|---|
| Python | 3.11 |
| PyTorch | 2.7.1+cu128 |
| SageAttention | 2.2.0 |
| vLLM | 0.9.2 |
| sgl-kernel | 0.1.8 |
🎬 480P 5s Video Test¶
Test Configuration:
- Model: Wan2.1-I2V-14B-480P-Lightx2v
- Parameters:
infer_steps=40,seed=42,enable_cfg=True
📊 Performance Comparison Table¶
| Configuration | Inference Time(s) | GPU Memory(GB) | Speedup | Video Effect |
|---|---|---|---|---|
| Wan2.1 Official | 366 | 71 | 1.0x | |
| FastVideo | 292 | 26 | 1.25x | |
| LightX2V_1 | 250 | 53 | 1.46x | |
| LightX2V_2 | 216 | 50 | 1.69x | |
| LightX2V_3 | 191 | 35 | 1.92x | |
| LightX2V_3-Distill | 14 | 35 | 🏆 26.14x | |
| LightX2V_4 | 107 | 35 | 3.42x |
🎬 720P 5s Video Test¶
Test Configuration:
- Model: Wan2.1-I2V-14B-720P-Lightx2v
- Parameters:
infer_steps=40,seed=1234,enable_cfg=True
📊 Performance Comparison Table¶
| Configuration | Inference Time(s) | GPU Memory(GB) | Speedup | Video Effect |
|---|---|---|---|---|
| Wan2.1 Official | 974 | 81 | 1.0x | |
| FastVideo | 914 | 40 | 1.07x | |
| LightX2V_1 | 807 | 65 | 1.21x | |
| LightX2V_2 | 751 | 57 | 1.30x | |
| LightX2V_3 | 671 | 43 | 1.45x | |
| LightX2V_3-Distill | 44 | 43 | 🏆 22.14x | |
| LightX2V_4 | 344 | 46 | 2.83x |
🖥️ RTX 4090 Environment (~24GB VRAM)¶
📋 Software Environment Configuration¶
| Component | Version |
|---|---|
| Python | 3.9.16 |
| PyTorch | 2.5.1+cu124 |
| SageAttention | 2.1.0 |
| vLLM | 0.6.6 |
| sgl-kernel | 0.0.5 |
| q8-kernels | 0.0.0 |
🎬 480P 5s Video Test¶
Test Configuration:
- Model: Wan2.1-I2V-14B-480P-Lightx2v
- Parameters:
infer_steps=40,seed=42,enable_cfg=True
📊 Performance Comparison Table¶
| Configuration | Inference Time(s) | GPU Memory(GB) | Speedup | Video Effect |
|---|---|---|---|---|
| Wan2GP(profile=3) | 779 | 20 | 1.0x | |
| LightX2V_5 | 738 | 16 | 1.06x | |
| LightX2V_5-Distill | 68 | 16 | 11.46x | |
| LightX2V_6 | 630 | 12 | 1.24x | |
| LightX2V_6-Distill | 63 | 12 | 🏆 12.37x |
🎬 720P 5s Video Test¶
Test Configuration:
- Model: Wan2.1-I2V-14B-720P-Lightx2v
- Parameters:
infer_steps=40,seed=1234,enable_cfg=True
📊 Performance Comparison Table¶
| Configuration | Inference Time(s) | GPU Memory(GB) | Speedup | Video Effect |
|---|---|---|---|---|
| Wan2GP(profile=3) | -- | OOM | -- | -- |
| LightX2V_5 | 2473 | 23 | -- | |
| LightX2V_5-Distill | 183 | 23 | -- | |
| LightX2V_6 | 2169 | 18 | -- | |
| LightX2V_6-Distill | 171 | 18 | -- |
📖 LightX2V Configurations¶
All LightX2V configurations use Wan2.1 I2V, SageAttention2, and 81 output frames. The JSON files in configs/bench specify inference steps, quantization, caching, and offload. Set computation precision through environment variables: use DTYPE=BF16 and set SENSITIVE_LAYER_DTYPE according to the table below.
| Benchmark | Configuration File | Sensitive Layer Precision |
|---|---|---|
| LightX2V_1 | wan_i2v.json | FP32 |
| LightX2V_2 | wan_i2v.json | BF16 |
| LightX2V_3 | wan_i2v_fp8.json | BF16 |
| LightX2V_3-Distill | wan_i2v_fp8_distill.json | BF16 |
| LightX2V_4 | wan_i2v_fp8_teacache.json | BF16 |
| LightX2V_5 | wan_i2v_fp8_offload.json | FP32 |
| LightX2V_5-Distill | wan_i2v_fp8_offload_distill.json | FP32 |
| LightX2V_6 | wan_i2v_fp8_offload.json | BF16 |
| LightX2V_6-Distill | wan_i2v_fp8_offload_distill.json | BF16 |
- Distill: 4-step DMD2 distillation with CFG disabled.
- FP8: uses Tiling VAE.
- Offload: block-level DiT CPU offload and T5 CPU offload.
- TeaCache: threshold of
0.2.
▶️ Running a Benchmark¶
Select the configuration file and sensitive layer precision for the benchmark. Set model_path to the base model directory for the chosen resolution. For FP8 configurations, replace dit_quantized_ckpt in the JSON with the DiT quantized weight file or directory compatible with fp8-sgl. Distilled configurations require the corresponding StepDistill-CfgDistill weights.
| Resolution | Size Arguments | Seed |
|---|---|---|
| 480P | --size 480 832 |
--seed 42 |
| 720P | --size 720 1280 |
--seed 1234 |
This example runs LightX2V_1 / 480P on H200:
lightx2v_path=path/to/LightX2V
model_path=path/to/Wan2.1-I2V-14B-480P-Lightx2v
export CUDA_VISIBLE_DEVICES=0
export DTYPE=BF16
export SENSITIVE_LAYER_DTYPE=FP32
source "${lightx2v_path}/scripts/base/base.sh"
python -m lightx2v.infer \
--model_cls wan2.1 \
--task i2v \
--model_path "${model_path}" \
--config_json "${lightx2v_path}/configs/bench/wan_i2v.json" \
--image_path "${lightx2v_path}/assets/inputs/imgs/img_2.jpg" \
--prompt "A close-up cinematic view of a person cooking in a warm,sunlit kitchen, using a wooden spatula to stir-fry a colorful mix of freshvegetables—carrots, broccoli, and bell peppers—in a black frying pan on amodern induction stove. The scene captures the glistening texture of thevegetables, steam gently rising, and subtle reflections on the stove surface.In the background, soft-focus jars, fruits, and a window with natural daylightcreate a cozy atmosphere. The hand motions are smooth and rhythmic, with a realisticsense of motion blur and lighting." \
--negative_prompt "镜头晃动，色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" \
--size 480 832 \
--seed 42 \
--save_result_path "${lightx2v_path}/save_results/lightx2v_1.mp4"
For another benchmark, set --config_json and SENSITIVE_LAYER_DTYPE according to the table, and select the corresponding weights, resolution, seed, and output path. All configurations use --model_cls wan2.1 --task i2v.
🔎 Comparison Implementations¶
| Implementation | Description |
|---|---|
| Wan2.1 Official | Original implementation from the Wan2.1 official repository |
| FastVideo | Based on the FastVideo official repository, using the SageAttention2 backend |
| Wan2GP(profile=3) | Based on the Wan2GP repository, using MMGP optimization. Profile 3 targets RTX 3090/4090 systems with at least 32GB RAM and 24GB VRAM, trading higher VRAM usage for lower system memory requirements. Uses quantized models: 480P model and 720P model |