How AI Talking Photo Works: Wan 2.2 vs Wav2Lip Explained (2026)

How AI Talking Photo Works: Wan 2.2 vs Wav2Lip

An "AI talking photo" turns a single still image into a video of that face speaking — with realistic lip sync, head motion, and expression. Two architectures dominate the space in 2026: Alibaba's Wan 2.2 and the long-established Wav2Lip lineage. Here's what each is good at and how to pick.

The Core Problem

Given a still photo and an audio track, generate a video where the face in the photo lip-syncs to the audio. Bonus points for natural blinks, head sway, and emotional expression that matches the audio's tone.

Wav2Lip — The Veteran

Wav2Lip was published in 2020 (arXiv:2008.10010) and remains the workhorse for lip-sync-only tasks. Its specialty is mouth-region replacement: it takes existing video and replaces the mouth area to sync with new audio. For pure lip-sync on existing video, it remains state of the art.

Strengths: Extremely fast inference, excellent lip accuracy, well-understood failure modes.

Limitations: Doesn't generate head motion or expression — works best on existing footage, not on still photos.

Wan 2.2 — The Generalist

Wan 2.2 is Alibaba Tongyi Lab's 2025 video diffusion model with character animation capabilities. Unlike Wav2Lip, it can take a single still photo and generate full-body video, including head motion, blinks, micro-expressions, and lip sync to a provided audio track.

Strengths: Generates realistic motion from a single image, handles full body when needed, produces emotional expression matched to audio tone.

Limitations: Slower inference (typically 30–90 seconds for a 10-second clip on H100), higher hardware cost, more variable output — sometimes needs a re-roll.

How Each Model Handles a Still-to-Video Job

Imagine a single front-facing photo of a person and a 10-second audio clip of them speaking.

• Wav2Lip alone: Cannot do this directly. It needs existing video to modify.
• Wan 2.2 alone: Generates the whole 10-second video from scratch — head motion, expression, lip sync.
• Hybrid pipeline: Some 2025 production stacks use Wan 2.2 for head motion + a Wav2Lip refinement pass on the mouth region. The hybrid often beats either alone on lip accuracy without sacrificing motion realism.

Identity Preservation

Both models lean on a face-embedding network for identity preservation. Production tools typically pair them with ArcFace or AdaFace embeddings to keep the generated frames identifiable as the original person. AdaFace shines on lower-quality source images.

When to Choose Which

• Wav2Lip: You already have video footage and need to overdub it (translation, subtitle replacement, dialogue replacement).
• Wan 2.2: You have only a still photo and want a full talking video. Or you need expressive emotional output.
• Hybrid: You need cinema-grade lip accuracy on still-to-video output and have the inference budget.

What FaceSwapAI Uses

FaceSwapAI's talking-photo feature uses Wan 2.2 by default and supports a Wav2Lip refinement pass for lip-critical content (translation, ADR, language localization). For most consumer use cases, Wan 2.2 alone is the right balance of quality and speed.

Hardware and Cost Snapshot

On an A100 (80 GB), a 10-second Wan 2.2 generation runs roughly 60–120 seconds. On H100, that drops to 25–45 seconds. Wav2Lip is closer to real-time on either GPU. For consumer browser tools, expect 1–2 minutes per 10-second clip end-to-end including queue time.

Limitations That Still Matter in 2026

• Both models perform worse on side-profile source photos (faces past ~45°).
• Lip sync on plosives (p, b, m) still occasionally falls behind audio by a few frames.
• Long clips (30 seconds+) accumulate temporal-coherence drift in pure single-image input — multi-frame anchoring helps.
• Languages with non-Latin phonemes (Mandarin tones, click consonants) need fine-tuned variants for best lip sync.

Bottom Line

Wav2Lip is the precision tool for video-to-video lip sync. Wan 2.2 is the canvas for image-to-video generation. Pick by your input format, not by hype. And if you're a creator just trying things out, the talking-photo demo on FaceSwapAI ships with Wan 2.2 ready to go — try it with one of your own photos and a 10-second voice memo before reading any more research papers.