Model Selection

Read this after the ad angle is clear.

Core Principle

Choose the model stack based on the production route, not on hype.

No serious UGC pipeline should assume one model does:

• avatar performance
• product photography
• proof-heavy B-roll
• localization
• variant generation

Current Recommendations

These are directional recommendations for v1. Always confirm actual availability and parameters with:

pica model search "<keyword>"
pica model info <model-id>

Route Mapping

1. AI avatar UGC

Use when:

• you need many variants fast
• you need localization
• you need script control and repeatability

Look for:

• dedicated avatar / lip-sync models
• performance-transfer models when you have driving video

Good fit:

• kling-avatar style workflows
• performance transfer with a source performance clip

Weak fit:

• pure text-to-video models with no reliable speech-performance control

2. Product-led or hybrid UGC

Use when:

• product proof matters more than face realism
• you need multiple short ad shots
• you want dynamic ad-style montage

Strong direction:

• Seedance 2.0 for multi-asset ad construction and short commercial sequences

Reality check:

• treat Seedance 2.0 as a high-potential ad model family, not a universal replacement
• verify current platform support, API status, and human-image restrictions before hard-coding it into automation

Pragmatic fallback:

• Seedance 1.5 pro or equivalent currently exposed commercial video models

3. Static assets and reference frames

Use image models for:

• creator portraits
• product shots
• first / end frames
• packaging and benefit callout stills

Strong direction:

• Seedream 5.0 Lite / newer Seedream family for instruction following, reference consistency, and clean ad stills

4. Premium cinematic ad footage

If the requirement is:

• strong overall realism
• native audio
• premium single-shot quality

Then consider high-end video families like Veo or top Kling variants.

But do not confuse that with "best for UGC". Premium realism is not the same as best direct-response performance.

Decision Table

Need	Preferred direction
Fast localized talking head	avatar / lip-sync model
Real performance transferred to avatar	performance transfer / driving-video model
Product-led short ad montage	Seedance-style ad video model
Clean stills and reference images	Seedream-style image model
Premium native-audio hero shot	Veo / top cinematic video family

Draft vs Final

For drafts:

• shortest duration
• lower resolution
• cheaper tier

For finals:

• only upscale or re-render the winning creative concept

Never spend premium budget discovering the hook.

What To Verify With model info

Before using any model, verify:

• aspect ratio support, especially 9:16
• duration limits
• native audio support
• image / audio / video input support
• reference image count
• first-frame / last-frame support
• whether captions or overlay text should be done in-model or in post