Model Details

This model is an NVFP4 W4A4 quantized version of Wan-AI/Wan2.2-I2V-A14B-Diffusers, generated using Intel AutoRound with round-to-nearest (RTN) quantization. Please follow the original model's license.

Quantized: both diffusion transformer experts, transformer and transformer_2.
Format: NVFP4 weights and activations, group size 16; exported as auto_round:llm_compressor.
Activations: dynamic quantization with calibrated static global scales (nv_fp4_with_static_gs).
Not quantized to NVFP4: text encoder (UMT5), VAE, and excluded transformer layers. The tokenizer and scheduler are unchanged.
Quantization: plain RTN, 0 tuning iterations, optimized RTN disabled; 32 OpenS2V calibration samples for activation scales. RTN does not mean calibration-free here.

How to Use

Environment

The evaluated setup used NVIDIA B200, Python 3.12.3, PyTorch 2.14.0, vLLM 0.33.0, Diffusers 0.40.0, FlashInfer 0.7.2.post1, NVIDIA CUTLASS DSL 4.8.0, and CUDA toolkit 12.8.93, with official vLLM-Omni commit 7465c0f52. Use a compatible environment with support for AutoRound NVFP4 checkpoints.

Inference used TRTLLM_ATTN attention and the FlashInfer NVFP4 backend (FlashInferCuteDslNvFp4LinearKernel in the evaluated runtime). Keep the exported quantization metadata unchanged. Other hardware and software combinations have not been validated by these results.

Image-to-Video Example

From the vLLM-Omni checkout above, run its official offline example. Set MODEL to the downloaded NVFP4 checkpoint directory or its Hugging Face repository ID.

curl -fL https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG -o input.jpg
MODEL=./wan22_i2v_nvfp4_rtn
CUDA_VISIBLE_DEVICES=0 DIFFUSION_ATTENTION_BACKEND=TRTLLM_ATTN \
python examples/offline_inference/image_to_video/image_to_video.py \
  --model "$MODEL" \
  --model-class-name WanImageToVideoPipeline \
  --image input.jpg \
  --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." \
  --negative-prompt "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" \
  --num-frames 81 --fps 16 --num-inference-steps 40 \
  --guidance-scale 3.5 --guidance-scale-high 3.5 \
  --boundary-ratio 0.9 --flow-shift 3.0 --seed 0 \
  --enforce-eager --init-timeout 7200 --stage-init-timeout 7200 \
  --output nvfp4.mp4

Without explicit dimensions, the example follows the input image's aspect ratio using max_area=480*832, with dimensions rounded down to multiples of 16. The quantitative evaluation below instead fixes dimensions to 832 × 480. The white-cat example is separate from the quantitative evaluation below.

White-Cat Video Comparison

Input image: the official white-cat surfing image used in the base model's example.

The comparison uses the full prompt and negative prompt above, automatic dimensions of 544 × 720, 81 frames, 40 inference steps, CFG 3.5/3.5, seed 0, and 16 fps, with boundary ratio 0.9 and flow shift 3.0. Both models use the same official vLLM-Omni example and environment; only the checkpoint differs. The BF16 reference is the same video used in the MXFP8 model card.

BF16 Reference	NVFP4 (This Model)

Download BF16 video	Download NVFP4 video

This individual showcase pair is not part of the 832 × 480 benchmark sets below.

Evaluation

VBench I2V Results

Less than 1 percentage point average absolute score drop versus BF16 on the five metrics below: 0.994 percentage points, calculated from unrounded scores with equal weight per metric. This is an absolute score difference on a 0–100 scale, not an official VBench aggregate score.

Evaluated with VBench 0.1.5 on the same 355 unique prompt/image pairs for BF16 and NVFP4: the union of the official i2v_subject (246 pairs) and i2v_background (109 pairs) subsets. Inputs use VBench's official cropped 16:9 reference images. Each model generated one video per pair.

Each metric uses only the pairs tagged for that dimension in the official I2V metadata. Scores are fractions; higher is better. The absolute drop in percentage points is 100 * (BF16 - NVFP4).

Metric	Videos per model	BF16	NVFP4	Absolute drop (percentage points)
i2v_subject	246	0.977584	0.961735	1.585
i2v_background	109	0.991074	0.982622	0.845
subject_consistency	246	0.952603	0.936307	1.630
background_consistency	109	0.969536	0.961778	0.776
motion_smoothness	246	0.984598	0.983227	0.137
Equal-weight mean	—	0.975079	0.965134	0.994

This is a one-video-per-pair subset evaluation, not the standard five-sample full VBench protocol; no VBench total score is reported.

Generation setting	Value
Resolution / frames / fps	832 × 480 / 81 / 16
Inference steps	40
Guidance scales	3.5 / 3.5, both experts
Seed	0, reset for each pair
Boundary ratio / flow shift	0.9 / 3.0
Negative prompt	Official Wan negative prompt, shown above
Input image resize	LANCZOS, via the official vLLM-Omni example
Runtime	Eager; no cache, CPU offload, or VAE tiling/slicing

Generation reused the unchanged official vLLM-Omni I2V example. Scoring called the official per-dimension functions with the metadata filtering described above, local=True, read_frame=False, and resolution="16-9". Standard image/frame preprocessing was unchanged. A narrowly scoped PyTorch safe-global alias for typing.OrderedDict was used for AMT checkpoint loading; weights-only loading was not disabled.

Generate the Model

The checkpoint was generated with AutoRound 0.17.0 from commit a85997e19. The command below preserves the recorded quantization settings, using portable model and output paths:

uv pip install git+https://github.com/intel/auto-round.git@a85997e19d97762172155c93c4dfc3fcc3976e71
auto-round \
  --model Wan-AI/Wan2.2-I2V-A14B-Diffusers \
  --scheme NVFP4 --algorithm rtn --disable_opt_rtn --iters 0 \
  --dataset opens2v --nsamples 32 --batch_size 1 \
  --calib_num_inference_steps 8 --num_inference_steps 40 \
  --guidance_scale 5.0 --generator_seed 42 --seed 42 \
  --device_map 0 --low_gpu_mem_usage \
  --diffusion_tuning_cache_size auto --disable_model_free \
  --format auto_round --output_dir wan22_i2v_nvfp4_rtn

The tuning-cache option was present in the recorded command but does not enable tuning when --iters 0. Calibration settings differ from evaluation settings; the 32 OpenS2V samples calibrate activation global scales, not the reported metrics.

Limitations

The largest measured reductions are in subject matching and subject consistency. Inspect subject identity, fine details, and temporal artifacts for the intended application. These five metrics do not fully measure aesthetics, prompt adherence, motion magnitude, or visual defects. Results cover a limited input subset and a single seed, without multi-seed confidence intervals; the below-one-percentage-point mean drop is a result for this run, not a universal quality guarantee.

This model inherits the base model's capabilities and limitations. Generated videos may contain artifacts, inaccurate content, or bias. Perform appropriate quality and safety testing before deployment, and comply with the original model's license and applicable requirements.