infinitetalk-shopping-avatar

InfiniteTalk Xiaosheng Shopping Guide Digital Human Sales Prompt Generation

Task Objective

• This skill is used to: generate structured prompts for InfiniteTalk project Xiaosheng Shopping Guide digital human sales videos, directly interfacing with the model inference process
• Capabilities include:
  • Generate storyboarded prompts adapted for InfiniteTalk Image-to-Video mode
  • Frame-level action timing description (5s/act, precise to the second level)
  • Audio synchronization matching (Suno music + chinese-wav2vec2-base encoding)
  • Consistency control (three major dimensions: character/visual/emotion)
  • Cross-tool integration (Suno, AI painting, InfiniteTalk)
• Trigger condition: the user needs to generate a Xiaosheng Shopping Guide digital human sales video and use the InfiniteTalk model for inference

Prerequisites

• Dependency description: this skill is based on the agent's natural language capabilities and requires no Python dependencies
• Environment preparation:
  • InfiniteTalk model environment (already installed, including the chinese-wav2vec2-base encoder)
  • TeaCache + int8 quantization configuration (already configured)
  • Character reference image (9:16 ratio, resolution ≥1080×1920)

Operation Steps

Standard Process

1. Prompt Generation

The prompt generation specialist is responsible for generating structured prompts adapted for InfiniteTalk:

Input Information：

• Scene type (product recommendation/price comparison/promotional activity, etc.)
• Product information
• Emotional tone
• Five-act emotional progression

Generation Process：

1. Read references/infinitetalk-parameters.md, confirm the core parameter configuration
2. Read references/prompt-structure.md, follow the prompt structure specification
3. Read references/scene-templates-infinitetalk.md, retrieve the scene template
4. Generate five-act prompts, each act includes:
   • Fixed character features (strictly follow the fixed description)
   • Action timing (frame-level description within 5s, precise to the second level)
   • Scene environment (background, foreground, props)
   • Audio matching (voice type + music style + audio alignment rules)
   • Lighting/atmosphere (color temperature, lighting type, emotional tone)
   • Technical constraints (composition ratio, depth of field, motion strength, etc.)

Output Format：

• Chinese prompt (can be directly used as thepromptparameter of the InfiniteTalk inference script)
• Parameter configuration table (duration, aspect_ratio, motion_strength, face_consistency, etc.)

2. Quality Inspection

The quality control specialist is responsible for inspecting prompt quality:

Inspection Dimensions：

1. Technical parameter matching degree：

   • Duration: 5s per act, total duration 25s
   • Resolution/ratio: 1080×1920 (9:16 vertical)
   • Action continuity: frame-level action descriptions within 5s are smooth
   • Character consistency: strictly match the fixed feature description

2. Character consistency：

   • Appearance features: face shape, eyes, hairstyle, skin tone, lip color
   • Clothing description: tops, shoes, accessories
   • Temperament characteristics: professional and approachable, hand posture

3. Audio compatibility：

   • Voice type: speech rate, tone, and emotion matching
   • Music style: matching the scene emotion
   • Audio alignment: movement rhythm synchronized with music/speech

4. Lighting and shadow feasibility：

   • Color temperature: 4000K-5500K range
   • Lighting type: side light, front light, top light, backlight
   • Lighting and shadow effects: light-dark contrast, depth of field, degree of blur

Verification results：

• Pass: hand over to the knowledge base operations engineer for archiving
• Fail: generate a problem list and feed it back to the prompt generation engineer for adjustment (maximum 2 times)

3. Knowledge base archiving

The knowledge base operations engineer is responsible for archiving prompts and updating the knowledge base:

Archived content：

• Character fixed feature template
• Five-act emotion-lighting mapping table
• Audio style-movement matching library
• Scene template library

Archive format：

• JSON format (convenient for program invocation)
• Includes complete metadata (scene type, emotional tone, parameter configuration)

4. Cross-tool integration

The cross-stage adaptation engineer is responsible for cross-tool integration and audio/image adaptation:

Audio integration：

1. Suno music generation:

   • Extract emotion keywords and generate music style descriptions
   • Export as 16kHz mono wav format
   • Ensure duration is 5s/act, strictly aligned with the prompts

2. chinese-wav2vec2-base encoding:

   • Use chinese-wav2vec2-base to perform feature encoding on the audio
   • Generate audio feature files

3. Salesperson voice generation:

   • Adjust speech rate/tone according to the emotion of each act
   • Strictly align with the music duration (5s/act)
   • Ensure lip synchronization

Image integration：

1. AI painting generation:

   • Use AI painting to generate the 'pre-action frame'
   • Crop to 9:16 ratio, resolution ≥1080×1920
   • Preserve the character's complete features (no occlusion)

2. Reference image adaptation:

   • As InfiniteTalk'sinit_imageparameter input
   • Ensure the character's fixed features are consistent with the prompts

InfiniteTalk inference integration：

1. Load model (basic configuration)
2. Single-act generation (taking the first act as an example)
3. Save video (5s/act)
4. Five-act stitching (total duration 25s)

Optional branches

• When the scene type is product recommendation: execute references/scene-templates-infinitetalk.md the product recommendation scene template in
• When the scene type is price comparison: execute the price comparison scene template
• When the scene type is promotional campaign: execute the promotional campaign scene template
• When cross-tool integration is needed: execute references/cross-tool-integration.md the integration solution in

Resource index

• Core parameter configuration: see references/infinitetalk-parameters.md(when to read: confirm the parameter configuration before generating prompts)
• Prompt structure specification: see references/prompt-structure.md(when to read: follow the structure specification when generating prompts)
• Scene template library: see references/scene-templates-infinitetalk.md(When to read: retrieve the template according to the scene type)
• Consistency control rules: see references/consistency-rules-infinitetalk.md(When to read: follow the rules during quality inspection)
• Cross-tool integration solution: see references/cross-tool-integration.md(When to read: refer to the solution during cross-tool integration)
• Complete example output: see assets/examples/sample-prompts-infinitetalk.md(When to read: refer to the complete example)

Notes

• Strictly follow the fixed character feature description; all five-act prompts use the same description
• Action timing must be precise to the second, ensuring actions are coherent within 5s
• Audio matching must be synchronized with the action rhythm, ensuring lip sync
• Lighting descriptions must be achievable, with color temperature within the 4000K-5500K range
• Technical constraints must match InfiniteTalk capabilities to ensure generation quality
• Fully leverage the agent's natural language capabilities and analytical reasoning abilities, avoiding writing scripts for simple tasks

Usage examples

Example 1: Product recommendation scene prompt generation

Function description: Generate five-act prompts for a product recommendation scene

Execution method: Prompt generator (agent natural language generation)

Key parameters：

• Scene type: Product recommendation
• Emotional tone: Enthusiastic and professional
• Five-act emotional progression: Enthusiastic mystery → Professional detail → Focused and vivid → Confident and powerful → Encouraging and resolute
• Technical parameters: 9:16 vertical screen, 5s/act, 1080×1920 resolution

Output: Five-act Chinese prompts, parameter configuration table

Example 2: Quality inspection

Function description: Inspect prompt quality

Execution method: Quality controller (agent analytical reasoning)

Inspection dimensions: Technical parameter matching, character consistency, audio matching, lighting achievability

Output: Quality inspection report (pass/fail), issue list (if failed)

Example 3: Cross-tool integration

Function description: Suno music → InfiniteTalk audio integration

Execution method: Cross-stage adapter (agent understanding + natural language guidance)

Key steps：

1. Suno generates 16kHz mono wav audio
2. chinese-wav2vec2-base encoding
3. Salesperson voice generation (speech rate/tone matching emotion)
4. InfiniteTalk inference integration

Output: Audio encoding file, video generation result

Responsibilities of the four major agent roles

Prompt generator

• Core responsibilities: Output InfiniteTalk-specific prompts according to the prompt structure
• Input: Scene type, product information, emotional tone
• Output: Act-by-act prompt text, parameter configuration table
• Key capabilities: Understand InfiniteTalk technical parameters, create frame-level action timing descriptions

Quality controller

• Core responsibilities: Inspect the compatibility between prompts and InfiniteTalk capabilities
• Inspection dimensions: Action coherence, character consistency, lighting achievability
• Output: Quality inspection report, prompt revision suggestions
• Key capabilities: Analytical reasoning, technical constraint matching assessment

Knowledge Base Operations Specialist

• Core responsibilities: Archive prompts and knowledge base updates
• Archived content: Character fixed feature templates, emotion-lighting mapping table, audio style-action matching library
• Output: Knowledge base retrieval interface (JSON format)
• Key capabilities: Structured archiving, metadata management

Cross-stage Adapter

• Core responsibilities: Cross-tool integration (Suno, AI drawing, InfiniteTalk)
• Integration steps: Audio adaptation, image adaptation, InfiniteTalk inference integration
• Output: Audio encoding files, video generation results
• Key capabilities: Cross-tool understanding, format adaptation, process integration