Prompt extension

Why the README recommends it

Cinematic labels (light, composition, camera) were part of 2.2’s data work. A three-word prompt under-uses that. Extension fills camera, lighting, and motion so the UMT5 encoder sees a dense caption.

DashScope API

• Set DASH_API_KEY. International site also needs DASH_API_URL=https://dashscope-intl.aliyuncs.com/api/v1.
• Default models: qwen-plus for T2V, qwen-vl-max for I2V.
• Override with --prompt_extend_model. Target language: --prompt_extend_target_lang zh|en.

DASH_API_KEY=your_key torchrun --nproc_per_node=8 generate.py --task t2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-T2V-A14B --dit_fsdp --t5_fsdp --ulysses_size 8 --prompt "Two anthropomorphic cats…" --use_prompt_extend --prompt_extend_method dashscope --prompt_extend_target_lang zh

Local Qwen

• T2V: Qwen/Qwen2.5-14B-Instruct, 7B, or 3B depending on leftover VRAM.
• I2V: Qwen/Qwen2.5-VL-7B-Instruct or 3B.
• --prompt_extend_method local_qwen and --prompt_extend_model with a HF id or local path.

FAQ

Does extension run every diffusion step?

No. Once, before sampling. Cost is one LLM call, not 50 DiT steps.

Can I skip it?

Yes. All basic README commands work without --use_prompt_extend. Quality is what you lose on short prompts.