Wuli Platform — AI Image & Video Generation#

Generate AI images and videos via the Wuli.art open platform API. Supports text-to-image, multi-reference image editing, text-to-video, first-frame image-to-video, first-last-frame video, auto-video, and sound control on supported video models across 25+ active models including Qwen Image, Seedream, GPT Image 2, Gemini image models, 通义万相, Happy Horse, 可灵, Seedance, and MiniMax Hailuo, with automatic uploads and no-watermark downloads when available.

When to Use#

• User wants to generate images from a text description
• User wants to edit or transform one or more existing images with AI
• User wants to create video from a text prompt
• User wants to animate a static image into a video
• User wants to generate a video from a first frame and last frame
• User wants to generate a video from reference images and/or a reference video
• User needs high-resolution (up to 4K) AI artwork
• User wants batch image generation
• User needs to choose between multiple AI models for best results

Setup#

bash

export WULI_API_TOKEN="wuli-your-token-here"

Get your token: log in to wuli.art, click the API entry at the bottom-left corner.

No additional dependencies — uses only Python 3 standard library.

Quick Start#

bash

# Generate an image (simplest usage)
python3 skill.py --action image-gen --prompt "a serene mountain lake at sunrise"

# Generate a video
python3 skill.py --action txt2video --prompt "waves crashing on a golden beach at sunset"

# First-last-frame video
python3 skill.py --action flf2video --prompt "camera pushes through the scene"   --image_path ./start.jpg --end_image_path ./end.jpg

# Auto-video from a reference video
python3 skill.py --action auto-video --prompt "preserve the motion, make it cinematic"   --video_path ./input.mp4 --model "通义万相 2.6"

Complete Command Reference#

bash

python3 skill.py --action <action> --prompt "description" [options]

Actions#

ActionDescriptionReference Inputsimage-genText to imageNoneimage-editEdit or transform one or more reference images--image_url or --image_path``txt2videoText to videoNoneimage2videoAnimate one image into a videoExactly one --image_*``flf2videoFirst-last-frame video generationStart frame via --image_*, end frame via --end_image_*``auto-videoAuto-video from reference images and/or videos--image_*, --video_*, or both

Parameters#

ParameterDefaultDescription--prompt_(required)_Generation prompt, max 2000 chars--modelauto-selectedModel name (see Model Selection Guide)--aspect_ratio1:1 (image) / 16:9 (video)Aspect ratio--resolution2K (image) / 720P (video)Output resolution. Model-dependent values include 2K, 3K, 4K, 480P, 720P, 768P, 1080P``--qualitybackend defaultImage quality for supported models. Values include low, medium, high, 低画质, 标准画质, 高画质``--n1Number of images to generate (1-4, image-gen only)--image_url—Reference image URL(s). Supports comma-separated multiple URLs--image_path—Local reference image path(s). Supports comma-separated multiple files--end_image_url—End-frame image URL for flf2video``--end_image_path—End-frame local image path for flf2video``--video_url—Reference video URL(s) for auto-video, comma-separated--video_path—Local reference video path(s) for auto-video, comma-separated--duration5Video length in seconds--negative_prompt—Exclude unwanted elements--soundbackend defaultEnable sound for supported video models. If omitted, backend default behavior is used--no-sound—Disable sound for supported video models--optimizetruePrompt optimization is enabled by default and recommended for most prompts--no-optimize—Disable prompt optimization when you need fully raw prompt behavior

Inline References in Prompt ( @imageN / @videoN)#

The open platform API accepts inline placeholders inside the prompt to bind a phrase to a specific reference asset:

• @image1, @image2, ... refer to the 1st, 2nd, ... entry of --image_path / --image_url (in the order given)

• @video1, @video2, ... refer to the 1st, 2nd, ... entry of --video_path / --video_url

• No space allowed between @ and the type, or between the type and the index: @image1 ✅, @image 1 ✗, @ image1 ✗

• Case-insensitive: @Image1 / @IMAGE1 also work

• Index out of range (e.g. @image5 when only 2 images are passed) → backend rejects with HTTP 400

• Placeholder is rewritten into the canonical <<<image_N:url>>> form before reaching any model:

  • 可灵 series (O1 / V3 Omni / V3 / V2.6 / V2.5 Turbo) — natively understands <<<image_N>>> as the Nth entry of image_list
  • 通义万相 2.6 R2V / Happy Horse 1.0 R2V — replaced with the model's internal characterN role token
  • Other models (Qwen Image / Seedream / Hailuo / Seedance) — the tag is stripped down to plain text image_N and embedded inline; no semantic binding, but human readable

Example:

bash

python3 skill.py --action auto-video   --prompt "把 @image1 中的人物放进 @video1 的场景,保留 @image1 的服装"   --image_path ./hero.jpg --video_path ./scene.mp4   --model "通义万相 2.6"

Input Rules#

• image-edit accepts one or more reference images
• image2video requires exactly one reference image
• flf2video requires exactly two images: start frame + end frame
• auto-video accepts reference images, reference videos, or both
• --sound / --no-sound only affect video tasks, and only take effect on models that support sound control
• Local files and remote URLs are auto-uploaded to Wuli OSS before submission

Aspect Ratios#

RatioUse Case1:1Square — social media posts, avatars, icons16:9Widescreen — videos, desktop wallpapers, presentations9:16Vertical — phone wallpapers, stories, reels4:3Classic — photos, prints3:2Photography — DSLR-style landscape shots2:1 / 1:2Wide or tall image crops on supported image models5:4 / 4:5Print and portrait image crops on supported image models21:9Ultra-wide — cinematic banners (image only on supported models)4:1 / 1:4Extreme panorama or vertical crops on supported Gemini image models

Model Selection Guide#

Image Models — Which to Choose#

ModelBest ForResolutionRef ImagesCost智能IMAGE 2Highest-control image generation/editing with quality tiers1K, 2K, 4K4quality-based全能图片 2Gemini Flash image generation/editing1K, 2K, 4K49-20 credits全能图片 ProGemini Pro image generation/editing1K, 2K, 4K44-5 credits通义万相 2.7Newest Wan image model, up to 9 ref images2K, 4K91 credit (2K) / 3 credits (4K)Qwen Image 2.0 _(default)_General purpose, fast, versatile2K, 4K41 creditQwen Image TurboQuick drafts, iterations2K, 4K41 creditSeedream 5.0 LiteHigher detail at lower cost than premium Seedream2K, 3K84 creditsSeedream 4.5Photorealism, high-fidelity detail2K, 4K84 creditsSeedream 4.0Photorealism with broad API-side support2K, 4K84 credits

Recommendations:

• Fastest & cheapest: Qwen Image Turbo
• Best all-rounder: Qwen Image 2.0
• Best quality control: 智能IMAGE 2 with --quality low|medium|high
• Best new image models: 智能IMAGE 2, 全能图片 2, 全能图片 Pro
• Best detail at mid tier: Seedream 5.0 Lite
• Best quality for photos: Seedream 4.5
• Most ref images (up to 9): 通义万相 2.7
• Need 4K: 通义万相 2.7, Qwen Image 2.0, Qwen Image Turbo, or Seedream 4.5

Video Models — Which to Choose#

ModelBest ForResolutionDurationKey Modes通义万相 2.2 Turbo _(default)_Quick videos, low cost720P5sTXT, FF通义万相 2.7Newest Wan video model with stronger narrative720P-1080P5-15s (AUTO 5-10)TXT, FF, FLF, AUTO通义万相 2.6 FlashFast image-to-video720P-1080P5-15sFF通义万相 2.6Best all-rounder with auto-video720P-1080P5-15sTXT, FF, AUTOHappy Horse 1.1Newest Happy Horse video model, lower 1080P cost720P-1080P5-15sTXT, FF, AUTOHappy Horse 1.0Long-duration TXT/FF/AUTO video with video-reference edit720P-1080P3-15sTXT, FF, AUTO可灵 3.0 OmniRichest multi-input video workflow720P-1080P5-15sTXT, FF, FLF, AUTO可灵 O1Premium omni video quality720P-1080P5-10sTXT, FF, FLF, AUTO可灵 3.0High-quality first-last-frame video720P-1080P5-15sTXT, FF, FLF可灵 2.61080P-focused Kling generation1080P5-10sTXT, FF, FLF可灵 2.5 TurboLower-cost Kling first-last-frame workflows1080P5-10sTXT, FF, FLFSeedance 1.5 ProMotion-heavy videos, up to 12s480P-720P5-12sTXT, FF, FLFSeedance 1.0 ProBroad resolution coverage480P-1080P5-10sTXT, FF, FLFMiniMax Hailuo 2.3Cinematic text/video generation768P-1080P6-10sTXT, FFMiniMax Hailuo 2.3 FastFaster image-to-video768P-1080P6-10sFF

Recommendations:

• Fastest & cheapest: 通义万相 2.2 Turbo
• Best all-rounder: 通义万相 2.7 or 通义万相 2.6
• Best multi-input / auto-video: 可灵 3.0 Omni or 通义万相 2.7
• Best Happy Horse model: Happy Horse 1.1 for image references, Happy Horse 1.0 when you need video-reference auto-video
• Best premium quality: Happy Horse 1.1, 可灵 O1, or 可灵 3.0
• Best for first-last-frame: 通义万相 2.7, 可灵 3.0, 可灵 2.6, or Seedance 1.5 Pro
• Best for 1080P-only workflows: 可灵 2.6 or 可灵 2.5 Turbo

Examples#

Text to Image#

bash

# Simple generation
python3 skill.py --action image-gen --prompt "anime girl with blue hair in a garden"

# Batch generate 4 images to pick the best
python3 skill.py --action image-gen --prompt "cyberpunk cityscape at night" --n 4

# Photorealistic with premium model
python3 skill.py --action image-gen --prompt "photorealistic mountain landscape, golden hour"   --model "Seedream 4.5" --resolution 4K --aspect_ratio 16:9

# Higher-detail model
python3 skill.py --action image-gen --prompt "luxury product photography, studio lighting"   --model "Seedream 5.0 Lite" --resolution 3K

# New quality-tier image model
python3 skill.py --action image-gen --prompt "editorial product shot, crisp reflective studio lighting"   --model "智能IMAGE 2" --resolution 2K --quality high

# Prompt optimization is enabled by default
python3 skill.py --action image-gen --prompt "a cat"

# Disable optimization only when you need exact raw prompting
python3 skill.py --action image-gen --prompt "a cat" --no-optimize

Image Editing#

bash

# Edit a local image
python3 skill.py --action image-edit --prompt "add sunglasses and a hat"   --image_path ./photo.jpg

# Edit with multiple reference images
python3 skill.py --action image-edit --prompt "merge these references into one unified brand illustration"   --image_path "./ref1.jpg,./ref2.jpg"

# Edit a remote image
python3 skill.py --action image-edit --prompt "change background to sunset beach"   --image_url "https://example.com/photo.jpg"

# Style transfer
python3 skill.py --action image-edit --prompt "convert to oil painting style"   --image_path ./landscape.jpg --model "Seedream 4.5"

Text to Video#

bash

# Simple video
python3 skill.py --action txt2video --prompt "waves crashing on a golden beach at sunset"

# Longer duration with higher quality
python3 skill.py --action txt2video --prompt "a cat playing piano in a jazz club"   --model "通义万相 2.6" --duration 10 --resolution 1080P

# Explicitly enable sound on a supported model
python3 skill.py --action txt2video --prompt "a singer performing on a neon stage"   --model "通义万相 2.6" --duration 10 --sound

# Cinematic quality
python3 skill.py --action txt2video --prompt "slow-motion rain drops on a window"   --model "可灵 O1" --resolution 1080P --aspect_ratio 16:9

# Disable sound when you want silent output or a lower-credit tier on some models
python3 skill.py --action txt2video --prompt "slow aerial shot over a mountain lake"   --model "可灵 3.0 Omni" --duration 10 --no-sound

Image to Video (Animate)#

bash

# Animate a landscape photo
python3 skill.py --action image2video --prompt "slow zoom in with gentle wind blowing"   --image_path ./landscape.jpg --duration 5

# Animate from URL with premium model
python3 skill.py --action image2video --prompt "character turns head and smiles"   --image_url "https://example.com/portrait.jpg" --model "可灵 O1"

First-Last-Frame Video#

bash

# Use start and end frame to control motion
python3 skill.py --action flf2video --prompt "a smooth cinematic transition between these two shots"   --image_path ./start.jpg --end_image_path ./end.jpg --model "可灵 3.0"

# Mix local and remote references
python3 skill.py --action flf2video --prompt "the flower blooms naturally from frame one to frame two"   --image_path ./flower_closed.jpg   --end_image_url "https://example.com/flower_open.jpg"

Auto-Video#

bash

# Auto-video from a reference video
python3 skill.py --action auto-video --prompt "keep the action, upgrade it to a cinematic sci-fi look"   --video_path ./input.mp4 --model "通义万相 2.6" --duration 10 --sound

# Auto-video from multiple reference images
python3 skill.py --action auto-video --prompt "turn these keyframes into a coherent product reveal video"   --image_path "./frame1.jpg,./frame2.jpg,./frame3.jpg"   --model "可灵 3.0 Omni" --resolution 1080P

# Auto-video with both images and video
python3 skill.py --action auto-video --prompt "preserve the motion, but restyle everything as an anime sequence"   --image_url "https://example.com/style_ref.jpg"   --video_url "https://example.com/source.mp4"   --model "可灵 O1"

Workflow#

text

1. (Optional) Upload reference images/videos
   --image_path ./photo.jpg             → auto-uploaded to cloud storage
   --image_url https://...              → auto-downloaded and re-uploaded
   --video_path ./clip.mp4              → auto-uploaded to cloud storage
   --image_path "./a.jpg,./b.jpg"       → uploads multiple image refs

2. Submit generation task
   → Returns recordId for tracking

3. Auto-poll for completion
   → Images: polls every 5s (up to 5 min)
   → Videos: polls every 10s (up to 20 min)

4. Auto-download results
   → Fetches no-watermark version when available
   → Saves to current directory
   → Auto-opens on macOS/Linux/Windows

Troubleshooting#

• "WULI_API_TOKEN not set": Run export WULI_API_TOKEN="wuli-your-token". Get your token from wuli.art bottom-left corner → API. The token is sent as Authorization: Bearer <token>.
• "HTTP 401": Token is invalid or expired. Regenerate it on the wuli.art platform.
• "HTTP 429": Rate limited. Wait a few seconds and retry.
• "Error code 2001": Insufficient credits. Top up at wuli.art.
• "REVIEW_FAILED": Content moderation rejected the prompt. Rephrase to avoid sensitive content.
• "TIMEOUT": Generation took too long. Try a faster model or shorter duration.
• "image2video requires exactly one reference image": Use only one --image_* input for image2video.
• "flf2video requires exactly two reference images": Pass a start frame with --image_* and an end frame with --end_image_*, or provide exactly two image references total.
• "auto-video requires at least one reference image or video": Provide --image_*, --video_*, or both.
• Media upload fails: Supported image formats are jpg, jpeg, png, webp. Supported video formats are mp4, mov, avi, webm.
• Sound flag seems ignored: --sound / --no-sound only affect video models that expose sound control. Some models or modes may keep backend default behavior.

Tips#

• Prompt optimization is enabled by default and is recommended for most prompts, especially short or vague ones.
• Use --no-optimize only when you need the model to follow your raw prompt as directly as possible.
• Start with the default models (Qwen Image 2.0 / 通义万相 2.2 Turbo) when you want the cheapest and simplest path.
• Generate multiple images with --n 4 and pick the best.
• Use --negative_prompt to exclude unwanted elements, e.g. --negative_prompt "blurry, low quality, watermark".
• For video, start with 5-second duration to preview, then re-generate at longer duration once you're happy with the style.
• If a video model supports sound control, --no-sound can be useful for silent exports and may reduce credits on some models.
• For flf2video, keep the start and end frames visually consistent for smoother motion.
• For auto-video, start with one video or 2-3 images before scaling up to more references.
• All results are auto-downloaded without watermarks when available.

For complete API documentation including current model tables and upload flow, see references/【呜哩Wuli】开放平台 API 文档.md.