Use when generating video clips with Chinese video models — text-to-video (文生视频), image-to-video (图生视频), first/last-frame and reference-to-video across 4 platforms: Bailian (Wan/PixVerse/Kling/Vidu/HappyHorse), Jimeng (doubao-seedance), MiniMax (Hailuo), Hunyuan (hy-video)
Use when generating video clips with Chinese video models — text-to-video (文生视频), image-to-video (图生视频), first/last-frame and reference-to-video across 4 platforms: Bailian (Wan/PixVerse/Kling/Vidu/HappyHorse), Jimeng (doubao-seedance), MiniMax (Hailuo), Hunyuan (hy-video)
Generate short video clips using Chinese video models across four providers — Alibaba Cloud Bailian (Wan/PixVerse/Kling/Vidu/HappyHorse), Volcengine Ark (Jimeng/即梦), MiniMax (海螺 AI), and Tencent Hunyuan (混元).
Four modes, auto-selected from inputs:
Mode
Inputs
Default model
t2v 文生视频
prompt only
wan2.7-t2v-2026-04-25 (Bailian)
i2v 图生视频
prompt + --image
wan2.6-i2v-flash (Bailian)
kf2v 首尾帧
prompt + --image + --last-frame
pixverse/pixverse-c1-kf2v (Bailian)
r2v 参考生视频
prompt + --ref (1-7 images)
pixverse/pixverse-c1-r2v (Bailian)
Video generation is asynchronous: submit → poll every 10s → download MP4. Result URLs expire after 24h, so the script always downloads immediately.
Local images: Wan/HappyHorse accept base64 data URIs directly; PixVerse/Kling/Vidu auto-upload to DashScope OSS (48h); Jimeng/Hunyuan use base64; MiniMax uploads via its file API.
When to Use This Skill
User asks to 生成视频 / 文生视频 / 图生视频 / 首尾帧 / 参考生视频
User names a Chinese video model: 万相/Wan, 爱诗/PixVerse, 可灵/Kling, Vidu, HappyHorse, 即梦/Jimeng, 海螺/MiniMax, 混元/Hunyuan
User needs B-roll, animated stills, character-consistent clips, or frame transitions
User asks about model pricing, features, or wants to compare models → open docs/models.html
Workflow
Step 0: Prompt Refinement (interactive)
Run for t2v/i2v unless skipped (see §0.4). Claude polishes the user's prompt before generation.
0.1 Analyze the raw input
Dimension
Check
Subject
Who/what? Appearance, action, expression
Scene
Where? Background, environment, atmosphere
Lighting
Time of day? Light quality? (golden hour, neon, soft diffused, backlit)
Provider: --provider {bailian,jimeng,minimax,hunyuan} or auto-detect from model name
Model: --model flag, or provider default for the mode
Parameters: duration, resolution, ratio from variant suggestion or user override
Step 2: Confirm and generate
Show the final command and confirm with the user. Run the script; it blocks until the task finishes and saves the MP4.
Step 3: Deliver
Report output path, file size, and generation time. Save to cwd if no path given.
Cost note: video APIs bill per second of output. Confirm with user for long/many clips.
Providers
Alibaba Bailian 百炼
One API key (DASHSCOPE_API_KEY) covers 5 model families. Third-party models (PixVerse/Kling/Vidu/HappyHorse) are cn region only. Models: Wan (t2v/i2v, up to 15s), PixVerse (all 4 modes, 1-15s), Kling (t2v/i2v/kf2v + r2v on omni), Vidu (q3: 1-16s with audio; q2: 1-10s), HappyHorse (t2v/i2v, 3-15s).
Volcengine Ark (Jimeng 即梦)
ARK_API_KEY via https://ark.cn-beijing.volces.com/api/v3. Seedance 2.0: t2v/i2v up to 15s/2K, with audio, lip-sync, and camera motion. Ratios: 16:9, 9:16, 1:1, 21:9.
MiniMax 海螺 AI
MINIMAX_API_KEY via https://api.minimax.chat. MiniMax-H3 (Hailuo 3.0): t2v/i2v, 4-15s, 768P/1080P (2K via direct API), 24fps with native stereo audio. Legacy video-01: 6s at 720P, prompt optimizer on by default (--no-prompt-optimizer to disable). Host override: MINIMAX_API_BASE (e.g. https://api.minimaxi.com CN / https://api.minimax.io intl).
Tencent Hunyuan 混元
HUNYUAN_API_KEY via TokenHub. hy-video-1.5: t2v/i2v (5-10s, 720P, supports --duration/--seed). Experimental i2v-only: yt-video-2.0, yt-video-fx, yt-video-humanactor. Flags --resolution, --ratio, --audio, --camera-motion are not yet supported.
Model Selection Guide
Use case
Model
Provider
Best quality t2v, multi-shot
wan2.7-t2v-2026-04-25
Bailian
Fast action / combat
pixverse/pixverse-c1-t2v
Bailian
Smart storyboard + audio
kling/kling-v3-video-generation
Bailian
Long clips up to 16s + audio
vidu/viduq3-pro_text2video
Bailian
Douyin/XHS short-video
doubao-seedance-2-0-260128
Jimeng
Smooth motion, natural physics
MiniMax-H3
MiniMax
Animate an image (default)
wan2.6-i2v-flash
Bailian
Transition between two frames
pixverse/pixverse-c1-kf2v
Bailian
Character/subject consistency
pixverse/pixverse-c1-r2v
Bailian
Cheap drafts
wanx2.1-t2v-turbo, happyhorse-1.0-t2v
Bailian
Chinese t2v/i2v on Tencent
hy-video-1.5
Hunyuan
Portrait animation (experimental)
yt-video-humanactor
Hunyuan
Run python scripts/generate_video.py --list-models for the full model catalog.
Model Comparison Page
When the user wants to compare models, browse pricing, filter by features, or pick a model:
Open docs/models.html in the browser. It's a self-contained static page — no server needed. Use open (macOS) or xdg-open (Linux):
open docs/models.html
The page supports filtering by provider, mode (文生/图生/首尾帧/参考生), and features (audio/camera/multi-shot), with pricing and capability comparison across all 24 models.
