LastKV-Wan-1.3B

480p causal text-to-video generation with LastKV, based on Wan2.1-T2V-1.3B.

FP32 generator weights from step 20,500. This is the EMA version; raw weights are also available.

Paper results

VBench scores (%, higher is better), from Table 8 of the paper.

Model	Quality	Semantic	Total
Transformer	83.1	76.5	81.8
LastKV	83.5	78.0	82.4

The paper evaluates step 20,000 EMA; this release contains step 20,500.

Full VBench breakdown (16 metrics)

Individual scores from the same paper EMA evaluation (%, higher is better).

Metric	Transformer	LastKV
Subject consistency	93.2	93.3
Background consistency	95.0	95.5
Temporal flickering	97.3	97.3
Motion smoothness	98.5	98.4
Dynamic degree	81.1	79.2
Aesthetic quality	61.6	62.2
Imaging quality	64.1	66.7
Object class	91.5	93.5
Multiple objects	72.8	74.6
Human action	95.6	96.6
Color	88.0	87.4
Spatial relationship	66.1	69.6
Scene	52.9	53.6
Appearance style	21.0	22.3
Temporal style	24.0	24.2
Overall consistency	25.8	25.8

Use the matching LastKV implementation with config.json and the Wan2.1 VAE/text encoder.

Project page · Paper