Wan2.2 I2V A14B, four-step 720p: native LightX2V FP8 and NVFP4

Community post-training conversions of the high-noise and low-noise 720p four-step checkpoints released by LightX2V on April 12, 2026. The upstream source card describes training on a 720p dataset and changes to low-noise training. This repository contains the converted DiT weights; use a matching high/low pair.

Format	Expert	File	Bytes
FP8	high	wan2.2_i2v_A14b_high_noise_lightx2v_4step_720p_260412_fp8.safetensors	14,540,662,920
FP8	low	wan2.2_i2v_A14b_low_noise_lightx2v_4step_720p_260412_fp8.safetensors	14,540,662,920
NVFP4	high	wan2.2_i2v_A14b_high_noise_lightx2v_4step_720p_260412_nvfp4.safetensors	8,383,804,680
NVFP4	low	wan2.2_i2v_A14b_low_noise_lightx2v_4step_720p_260412_nvfp4.safetensors	8,383,804,680

The four files total 45,848,935,200 bytes (45.85 GB). Exact SHA256 hashes, source identities and conversion settings are in conversion-720p-260412.json.

Conversion and compatibility

Source repository: lightx2v/Wan2.2-Distill-Models, revision db93455b9e85c4d8a3ff9297fcfa189d213cfe29. The source files contain FP32 weights on disk. Conversion targets native LightX2V at 9bf4d39ac048a0d922b9a2139da9ed795f78f4e8.

• FP8: E4M3 weights with per-output-channel FP32 scales, using qtorch nearest rounding. The native block attention and feed-forward layers (400 matrices per expert) are quantized; remaining floating weights are BF16.
• NVFP4: packed E2M1 weights with E4M3 block scales, plus FP32 input_global_scale and alpha fields. Weights enter the quantizer as BF16. Each expert used its own activation calibration over four completed BF16 reference requests with task-specific LoRAs active at runtime (no LoRA weights were merged). This is a small calibration set. Remaining floating weights are BF16.
• No LoRA is merged. No additional training or attention-sparsity training was performed. These files do not recreate the separately trained LightWan Sparse checkpoints.
• These are native LightX2V layouts. Compatibility with ComfyUI loaders or generic Diffusers loading has not been established. The FP8 files are not NVIDIA's separate NVFP8 format.

Runtime setup

Download both files for one precision. Supply the Wan2.2 model configuration, UMT5 text encoder/tokenizer and Wan VAE separately; this repository is not a complete pipeline. NVFP4 inference was tested on an RTX PRO 6000 Blackwell (SM120) using the matching LightX2V CUDA kernel, PyTorch 2.11.0 + CUDA 13.0. Use a compatible compiled native kernel; hardware support is not implied for older GPUs.

In the LightX2V Wan2.2 I2V config, set:

{
  "dit_quantized": true,
  "dit_quant_scheme": "nvfp4",
  "high_noise_quantized_ckpt": "/models/wan2.2_i2v_A14b_high_noise_lightx2v_4step_720p_260412_nvfp4.safetensors",
  "low_noise_quantized_ckpt": "/models/wan2.2_i2v_A14b_low_noise_lightx2v_4step_720p_260412_nvfp4.safetensors",
  "infer_steps": 4,
  "distill_method": "dmd2",
  "enable_cfg": false,
  "sample_guide_scale": [1.0, 1.0],
  "sample_shift": 5.0,
  "boundary_step_index": 2,
  "denoising_step_list": [1000, 750, 500, 250]
}

This is a partial configuration, not a standalone launch file. For FP8, the tested configuration used dit_quant_scheme: fp8-triton and the corresponding FP8 file paths. The tested checkout included a local Triton fix for masked tail loads of scales/bias and zero-filled ragged activation loads; unpatched fp8-triton at the pinned upstream commit is not the validated runtime. Runtime image size, frames, attention backend and output FPS are configured separately. Start with dense attention when assessing conversion quality; dynamic sparse attention changes inference behavior independently of the checkpoint format.

Validation and limits

Both converted pairs completed GPU image-to-video inference on the RTX PRO 6000 Blackwell in a prepared LightX2V checkout with local single-GPU, large-seed and dynamic-LoRA fixes (plus the FP8 kernel fix above). These runtime patches are not included in the checkpoint files. This establishes loading and execution for the tested setup, not equivalent quality across all inputs, kernels or production workloads. Quantization and the limited NVFP4 calibration can affect results. Performance depends on attention, kernels, resolution, frame count, model residency and warm state; no universal speed claim is made here.

Input photos, generated videos, prompts, LoRAs and calibration sample records are not included. The model files contain only converted base weights and numeric quantization metadata.

Attribution and license

Original model: Wan-AI/Wan2.2-I2V-A14B. Four-step 720p source weights: LightX2V / ModelTC. This is a community conversion, not an official upstream release. The upstream model cards declare Apache License 2.0; see LICENSE and NOTICE. The changes in this distribution are weight quantization and the associated runtime scales described above.