Text-to-video

Text-to-video in the open 2.2 line is two recipes: the cinematic MoE T2V-A14B (480P/720P, ~80 GB) and the hybrid TI2V-5B (720P@24 fps on a 24 GB card with offload). Both decode silent frames.

t2v-A14Bti2v-5B720P
On this page 6
Text-to-video pipeline from prompt through UMT5 and DiT MoE to VAE decode
T2V path: prompt → UMT5-XXL → Wan2.2 DiT → VAE decode.

Two official T2V entries

Task flagModelTypical sizeVRAM (official)Notes
t2v-A14BWan2.2-T2V-A14B1280×720 or 480P≥ 80 GB single GPUMoE high/low-noise experts
ti2v-5BWan2.2-TI2V-5B1280×704 or 704×1280≥ 24 GB with offload720P @ 24 fps; hybrid T2V+I2V

Official single-GPU commands

T2V-A14B (80 GB class)
python generate.py --task t2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-T2V-A14B --offload_model True --convert_model_dtype --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
TI2V-5B (24 GB / 4090 class)
python generate.py --task ti2v-5B --size 1280*704 --ckpt_dir ./Wan2.2-TI2V-5B --offload_model True --convert_model_dtype --t5_cpu --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage"

Multi-GPU uses torchrun with --dit_fsdp --t5_fsdp --ulysses_size. The README example for eight GPUs is the supported path when a single 80 GB card is not available.

What the stack does

  • Text encoder: UMT5-XXL. Offload it with --t5_cpu if VRAM is tight.
  • Backbone: flow-matching DiT. A14B switches experts by SNR; TI2V-5B is dense.
  • Latents: Wan-VAE (2.1-style 4×8×8 on A14B; 2.2 4×16×16 on TI2V-5B).
  • Prompt extension is recommended; see the prompt-extension wiki.

Quality claims to take from the source

The README positions Wan2.2 as topping Wan-Bench 2.0 versus several closed models, with more motion/semantic/aesthetic data than Wan2.1. Treat third-party “SOTA” tables as marketing unless they cite the same bench and the same 2.2 checkpoint (A14B vs 5B vs a quantized community pack).

FAQ

Which size should I pass?
A14B uses 1280*720 (or a 480P pair). TI2V-5B 720P is 1280*704 or 704*1280 — not 1280*720. That difference is in the official README.
How long is a clip?
The public recipes are short cinematic shots (on the order of 5 seconds). S2V duration follows the audio track instead.
Diffusers?
Yes. Wan-AI/Wan2.2-T2V-A14B-Diffusers and Wan2.2-TI2V-5B-Diffusers are listed in the July 28, 2025 news block.

Primary sources

Checked against Wan-Video/Wan2.2 README · Aug 28, 2026

© 2026 wan2.video