A mixed INT4/INT8 OpenVINO conversion of the Wan2.2 TI2V 5B Diffusers model, calibrated with real transformer inputs. The single checkpoint supports image-to-video generation. This export currently uses a fixed transformer shape and is validated at 704 × 544 pixels for 21 frames.
Python 3.11 is recommended. Install the pinned runtime dependencies:
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
Generate a clip:
python infer.py \
--image path/to/input.jpg \
--prompt "A cinematic shot of a red fox walking through a snowy forest" \
--negative-prompt "blurry, distorted, subtitles" \
--output result.mp4 \
--steps 4 \
--seed 42
The exported transformer has static inputs for 704 × 544 and 21 frames. infer.py rejects other resolution/frame combinations so they do not fail later inside OpenVINO. The sample video uses 4 denoising steps and is encoded at 16 fps.
NNCF 3.4.0 compressed the transformer with symmetric INT4 weights, group size 64, ratio 0.9, and AWQ using eight captures from the official Diffusers pipeline. NNCF kept 10% of ratio-defining parameters in per-channel symmetric INT8 as a backup. The text encoder and both VAE components are exported as FP16 OpenVINO IR. The transformer IR is about 2.8 GB; the full IR package is about 15 GB.
The included quantization_config.json records the recipe and static input shapes. This is weight-only quantization; activations are not quantized.
The package was run end to end with OpenVINO 2026.4.0 on CPU, using the upstream example image, official sample prompt, 4 denoising steps, guidance 5.0, and seed 42. The generated 704 × 544, 21-frame clip completed in 893.6 seconds on a 30-vCPU Lambda host. The first frame was visually close to the BF16 CUDA reference generated with the same prompt, image, and seed. The included sample_i2v.mp4 and sample_first_frame.png are from that OpenVINO run.
Against the first frame of the BF16 reference, the quantized sample had mean absolute pixel error 0.383/255, RMSE 0.674/255, and 64.4% exactly matching channel values. This is one deterministic example and should not be treated as a broad quality evaluation.
OpenVINO's CPU plugin performed the validation; NVIDIA A100 is not an OpenVINO GPU target. Runtime depends on CPU hardware and thread configuration.
The packaged runner also passed end-to-end on an NVIDIA A100-SXM4-40GB with OpenVINO 2026.4.0 and the OpenVINO NVIDIA plugin. The OpenVINO text encoder and INT4/AWQ transformer ran on NVIDIA; the unquantized VAE ran through the upstream BF16 PyTorch implementation on CUDA. The upstream VAE weights are downloaded from the Apache-2.0 base model on first use.
Use the tested hybrid path with:
python infer.py \
--image path/to/input.jpg \
--prompt "A white cat wearing sunglasses rides a surfboard across turquoise water" \
--output result.mp4 \
--steps 4 \
--seed 42 \
--device NVIDIA \
--vae-backend cuda
Two 704 × 544, 21-frame clips passed visual inspection at four steps and guidance 5.0: the official white-cat beach prompt (seed 42) and a second tracking prompt (seed 43). The saved outputs are the beach sample and the tracking sample. On the seed-42 run, the NVIDIA OpenVINO transformer plus CUDA VAE output had MAE 24.50/255 and RMSE 37.98/255 against the BF16 CUDA video reference; the existing CPU OpenVINO package sample measured 24.36/255 and 37.78/255 against that same reference. First-step transformer output matched the source BF16 transformer at cosine 0.9989 / relative RMSE 0.0477. The text encoder, after the pipeline's padding normalization, matched at cosine 0.9986 / relative RMSE 0.0530.
The OpenVINO VAE IRs do not currently produce correct output through the NVIDIA plugin. The --vae-backend cuda option is required for this tested A100 path. The OpenVINO VAE path remains available for the CPU workflow above; Intel GPU execution has not been validated.
The OpenVINO package compiled all four neural components on Intel GPU and decoded a 21-frame 544×704 clip in 153.27 seconds on Arc B390. Its late frames were corrupted. A separate Wan Turbo Vulkan recipe completed the matched clip in 25.21 seconds with tiled TAEHV, retaining the subject and colors through the last frame. Both measurements used the B390, Windows 11, and driver 32.0.101.9033. Iris Xe has not been measured.
Base model: Wan-AI/Wan2.2-TI2V-5B-Diffusers. The upstream model is licensed under Apache 2.0. This conversion is provided under the same license; see LICENSE.txt. See the upstream model card for its model use terms and citations.
The conversion uses the official OpenVINO Wan2.2 notebook as the pipeline/export reference, with NNCF weight compression applied to the transformer.
text_encoder.bin 11 GBtransformer.bin 2.9 GBvae_decoder.bin 1.1 GBvae_encoder.bin 296 MB