bailian-multimodal-skills（百联多式联运技能）

Bailian Multimodal Skills

Generate images, audio, video, and transcribe speech using Aliyun Bailian (Qwen/Wan/PixVerse/Kling/CosyVoice) models.

Features

• Image Generation: z-image-turbo, wan2.6-t2i, wan2.7-image-pro
• Image Editing: wan2.7-image-pro
• Video Editing: wan2.7-videoedit
• ASR (Speech-to-Text): qwen3-asr-flash
• TTS (Text-to-Speech): qwen3-tts-flash
• Text-to-Video: wan2.7-t2v, wan2.6-t2v, pixverse/pixverse-v5.6-t2v, kling/kling-v3-video-generation
• Image-to-Video: wan2.7-i2v, wan2.6-i2v-flash, wan2.6-i2v, pixverse/pixverse-v5.6-it2v, kling/kling-v3-video-generation
• Reference-to-Video: wan2.6-r2v-flash, wan2.6-r2v, pixverse/pixverse-v5.6-r2v

Usage

1. Image Generation

Generate images from text.

uv run {baseDir}/scripts/run_multimodal.py --mode image --model z-image-turbo --prompt "A futuristic city" --output "city.png"

Models: z-image-turbo, wan2.6-t2i, wan2.7-image-pro

wan2.7-image-pro supports multi-image sequential generation and 2K size:

uv run {baseDir}/scripts/run_multimodal.py --mode image --model wan2.7-image-pro --prompt "电影感组图，记录同一只流浪橘猫的四季" --n 4 --enable-sequential --size 2K --output "cat_seasons.png"

Options (wan2.7-image-pro): --n (number of images), --enable-sequential (keep character consistency across images), --size (e.g., 2K, 1024*1024), --watermark

2. Image Editing

Edit images with text instructions. Supports multiple input images.

uv run {baseDir}/scripts/run_multimodal.py --mode image-edit --model wan2.7-image-pro --input-images "car.png" "graffiti.png" --prompt "把图2的涂鸦喷绘在图1的汽车上" --output "edited.png"

Options: --input-images (required, one or more image URLs/local paths), --prompt (required, editing instruction), --size (default: 2K), --n, --watermark, --thinking-mode (enable deeper reasoning for complex edits)

3. ASR (Speech Recognition)

Transcribe audio files or URLs to text.

uv run {baseDir}/scripts/run_multimodal.py --mode asr --model qwen3-asr-flash --input-audio "https://example.com/audio.mp3"

4. TTS (Speech Synthesis)

Convert text to speech.

uv run {baseDir}/scripts/run_multimodal.py --mode tts --model qwen3-tts-flash --text "Hello world" --output "hello.wav"

5. Text-to-Video (T2V)

Generate video from text prompt. Async task with auto-polling.

wan2.7-t2v (new protocol — uses resolution + ratio instead of size, no shot_type):

uv run {baseDir}/scripts/run_multimodal.py --mode t2v --model wan2.7-t2v --prompt "一只小猫在月光下奔跑" --resolution 1080P --ratio 16:9 --duration 10 --output "cat.mp4"

Options (wan2.7-t2v): --resolution (720P/1080P, default 1080P), --ratio (16:9/9:16/1:1, default 16:9), --duration, --prompt-extend/--no-prompt-extend, --negative-prompt, --audio-url, --audio/--no-audio, --watermark, --seed

Note: --shot-type is not supported for wan2.7-t2v. Use natural language in --prompt to describe shot structure (e.g., "生成多镜头视频" or timestamp-based descriptions).

wan2.6-t2v and earlier (legacy protocol — uses size and shot-type):

uv run {baseDir}/scripts/run_multimodal.py --mode t2v --model wan2.6-t2v --prompt "一只小猫在月光下奔跑" --duration 10 --size "1280*720" --output "cat.mp4"

Models: wan2.6-t2v, pixverse/pixverse-v5.6-t2v, kling/kling-v3-video-generation

Options: --size (e.g., 1280*720, 1920*1080), --duration, --prompt-extend/--no-prompt-extend, --shot-type single|multi, --negative-prompt, --audio-url, --audio/--no-audio, --watermark, --seed, --quality-mode std|pro

5.5 Video Editing (videoedit)

Edit an existing video with text instructions (style transfer, content modification, etc.).

uv run {baseDir}/scripts/run_multimodal.py --mode videoedit --model wan2.7-videoedit \
  --video-url "https://example.com/input.mp4" \
  --prompt "将整个画面转换为黏土风格" \
  --resolution 1080P \
  --output "edited.mp4"

With reference images (for appearance/style guidance):

uv run {baseDir}/scripts/run_multimodal.py --mode videoedit --model wan2.7-videoedit \
  --video-url "clip.mp4" \
  --prompt "为人物换上参考图里的服装" \
  --ref-images "outfit.png" \
  --resolution 1080P \
  --output "result.mp4"

Models: wan2.7-videoedit

Options:

• --video-url (required): input video URL or local file path
• --prompt (optional): editing instruction (up to 5000 chars)
• --negative-prompt (optional): content to avoid
• --ref-images (optional): one or more reference image URLs/local paths
• --resolution (720P/1080P, default 1080P)
• --ratio (16:9/9:16/1:1; omit to preserve input video's ratio)
• --prompt-extend/--no-prompt-extend (default: enabled)
• --watermark, --seed

Note: Billing = input video duration + output video duration (seconds).

6. Image-to-Video (I2V)

Generate video from a reference image (first frame).

wan2.7-i2v (new protocol — uses resolution + ratio):

uv run {baseDir}/scripts/run_multimodal.py --mode i2v --model wan2.7-i2v --img-url "photo.png" --prompt "人物微微歪头，自然眨眼，浅笑" --resolution 1080P --ratio 9:16 --duration 10 --output "result.mp4"

Options (wan2.7-i2v): --img-url (required), --prompt, --resolution (720P/1080P, default 1080P), --ratio (16:9/9:16/1:1), --duration, --prompt-extend/--no-prompt-extend, --negative-prompt, --audio-url, --audio/--no-audio, --watermark, --seed

wan2.6 and other models:

uv run {baseDir}/scripts/run_multimodal.py --mode i2v --model wan2.6-i2v-flash --img-url "https://example.com/cat.png" --prompt "A cat running" --resolution 720P --duration 5 --output "cat_run.mp4"

Models: wan2.6-i2v-flash, wan2.6-i2v, pixverse/pixverse-v5.6-it2v, kling/kling-v3-video-generation

Options: --img-url (required, image URL, base64, or local file path), --prompt, --resolution (480P/720P/1080P), --duration, --prompt-extend/--no-prompt-extend, --shot-type single|multi, --negative-prompt, --audio-url, --audio/--no-audio, --watermark, --seed, --quality-mode std|pro

7. Reference-to-Video (R2V)

Generate video with character/object references (images or videos as actors).

uv run {baseDir}/scripts/run_multimodal.py --mode r2v --model wan2.6-r2v-flash --prompt "character1 在公园里散步" --reference-urls "https://example.com/person.png" --size "1280*720" --duration 5 --output "walk.mp4"

Multi-character example:

uv run {baseDir}/scripts/run_multimodal.py --mode r2v --model wan2.6-r2v-flash --prompt "character1 对 character2 说你好" --reference-urls "https://example.com/role1.mp4" "https://example.com/role2.png" --shot-type multi --output "dialog.mp4"

Models: wan2.6-r2v-flash, wan2.6-r2v, pixverse/pixverse-v5.6-r2v

Options: --reference-urls (required, space-separated, up to 5, supports local file paths), --prompt (required, use character1/character2 to map references), --size, --duration, --shot-type single|multi, --negative-prompt, --no-audio, --watermark, --seed

Notes

• 本地文件处理：脚本自动处理本地路径。wan2.7-videoedit 模式会将本地视频/图片上传到 DashScope OSS 获取 oss:// URL；其他模式（image/i2v/r2v 等）使用 SDK 自带的 file:// URL 传入。
• PixVerse 与 Kling 模型以 URL 方式提交媒体；脚本已兼容本地路径输入，但若百炼后端对某些模型拒绝 file://，请改用可公网访问的 HTTP/HTTPS URL。

Configuration

API Key 按以下优先级读取：

1. 命令行参数 --api-key
2. 环境变量 DASHSCOPE_API_KEY
3. 配置文件 ~/.config/bailian-multimodal/api_key.txt

# 方式一：环境变量
export DASHSCOPE_API_KEY="sk-..."

# 方式二：配置文件
mkdir -p ~/.config/bailian-multimodal
echo "sk-..." > ~/.config/bailian-multimodal/api_key.txt