ALIBABA WAN-AI • Wan2.1-T2V-1.3B • APACHE-2.0 • VBENCH SOTA 1.3B

Wan2.1-T2V-1.3B Consumer-GPU Video Studio

Open-source 1.3B Flow-Matching Text-to-Video foundation model requiring just 8.19 GB VRAM (--offload_model True --t5_cpu), featuring causal 3D Wan-VAE spatiotemporal compression and native English & Chinese visual text generation at 832×480 (16 fps).

📦 Wan-AI/Wan2.1-T2V-1.3B 🧨 Diffusers Weights 💻 Official GitHub
Minimum Consumer VRAM
8.19 GB
Runs on RTX 4060 / 3060 / 4090 (~4m for 5s 480P)
Native 480P Geometry
832×480 / 81f
16 fps (5.06s) • Landscape 16:9 & Vertical 9:16
Recommended Flow Shift
shift 8 – 12
sample_guide_scale = 6.0 (UniPC Flow Matching)
Visual Typography
EN + 中文
First open T2V model rendering bilingual scene text

🚀 Wan2.1 Text-to-Video Live Cloud Generator

Generate 480P videos in your browser via @gradio/client (routes across Wan-AI/Wan2.1 async cloud & fast ZeroGPU endpoints) and sync parameters straight to your Diffusers / CLI script.

🏎️ Sunrise Coastal Supercar 🏮 Cyberpunk 中文 Neon Sign ☕ EN Chalkboard Café Typography 🌊 Ink-Wash Dragon Calligraphy
Ready • Select a preset or enter EN/中文 prompt.

🖥️ Active Video Monitor

Displays live cloud output or instant local Wan showcase clips.

Wan Showcase — Rainy Alley Hypercar Idle 832×480 • 16fps
1. Rainy Alley Hypercar 2. Aerial Chopper Searchlight 3. Open Bridge Rain POV 4. Airborne Bridge Stunt

🧮 Causal 3D Wan-VAE & Consumer GPU Fitment Profiler

Wan-VAE compresses spatiotemporal video pixels [3, F, H, W] by 4×8×8 into causal chunks [16, (F-1)/4 + 1, H/8, W/8], enabling unlimited-length 1080P feature-chunk decoding without temporal boundary flicker.

📐 Architecture Breakdown (Wan2.1-T2V-1.3B)

Why 1.3B punches above 5B–14B baselines on VBench:

1. Text Encoder:  UMT5-XXL (Multilingual EN + 中文 + 100+ langs)
                  → Offloadable to System RAM via --t5_cpu!
2. Diffusion DiT: 1.34B Flow-Matching Transformer (BF16 = ~2.68 GB)
                  • Full spatiotemporal 3D self-attention + cross-attn
                  • UniPC / Flow-DPMSolver++ (flow_shift = 8.0..12.0)
3. Causal 3D VAE: Wan-VAE (torch.float32 recommended for crisp decode)
                  • Stride: 4x temporal × 8x height × 8x width (16 ch)
                  • 81 × 480 × 832 RGB → [16, 21, 60, 104] latents
                  • Only 131,040 latent tokens per 5-second video!

🧨 Hugging Face diffusers Pipeline (Wan-AI/Wan2.1-T2V-1.3B-Diffusers)

Live-synchronized with your prompt, resolution, flow_shift, and guidance_scale in Tab 1:


      

      

        
💻 Official Wan2.1/generate.py 8.19 GB Consumer-GPU CLI

        
Exact low-VRAM command with --offload_model True --t5_cpu: