Animating a Still Portrait into a Talking Head with Lip Sync Tools

You feed a single JPEG to an AI model and get back a 5-second video of that person speaking your script. The lip sync is convincing for 2 seconds, then the jaw drifts left, the skin texture softens to plastic, and by frame 80 the person on screen is no longer the person in your photo. That sequence is the uncanny edge of image-to-video portrait talking head lip sync generation today. This page gives you the technical requirements to avoid the worst failure modes, the lip-sync tool landscape with exact vendor-documented limits, and the specific uncanny-valley failure modes you must watch for.

The source image determines whether the output is usable or a horror show. The model holds a single frame as its only reference. If that frame is ambiguous, the model fills the gaps from its training data and that is where identity shift happens.

Use a source image with even front-on lighting across the whole face. A shadow across one cheekbone, a highlight on one eye: the model interprets those as features to move, and they drift during speech. Use high resolution on the face. Runway Gen-4 accepts up to 4K source images (documented 2025-12-18). Kling 2.5 accepts JPEG, PNG or WebP up to 10 MB and recommends 16:9, 9:16, 1:1, 4:3 or 3:4 aspect ratios (documented 2025-12-18). A 720p face cropped from a group shot gives the model roughly 100 pixels across the mouth to work with. The model will hallucinate detail. Feed it a 2K headshot.

A closed mouth is preferred. An open-mouth source forces the model to decide whether the teeth and tongue are part of the face or a separate object. The model usually gets this wrong, producing teeth that flicker between frames or a tongue that slides sideways. Closed mouth gives the model a single, stable shape to animate.

The shoulder boundary is a frequent problem. If your frame includes shoulders, the arm-to-torso edge sits at a point where the model most commonly produces limb mutation. Crop just below the collar to remove that failure vector.

Every model listed here expects a source image plus a text prompt that describes the motion. None of them accept an audio file as input for lip sync. The motion is controlled by the text prompt describing what the person says or does, not by an external waveform.

Veo 3 (Google DeepMind, documentation checked 2025-12-18) accepts JPEG, PNG or WebP up to 4K. Native clip length is user-selectable at 4, 6 or 8 seconds at 720p or 1080p. It generates audio alongside the video but exposes no motion controls beyond the text prompt. No start or end frame, no motion strength slider, no camera control, no region masks. Access is through VideoFX or the Vertex AI API.

Kling 2.5 (Kuaishou, documented 2025-12-18) accepts JPEG, PNG or WebP up to 10 MB. Native clip length is 5 or 10 seconds at 1080p. It has start frame, optional end frame, motion brush for region-specific animation, and camera control presets for pan, tilt, zoom and roll. No audio generation. Access via the Kling web app or API.

Runway Gen-4 (documented 2025-12-18) accepts JPEG or PNG up to 4K source. Native clip length is 5 or 10 seconds at 720p or 1080p. Motion controls include a motion strength slider (0-10), camera control (pan, tilt, zoom, dolly, crane, roll), and start frame image input. No end frame, no region masks, no audio generation.

Sora (OpenAI, documented 2025-12-18) accepts JPEG, PNG or WebP up to 20 MB. Native clip length is 4, 8 or 12 seconds at up to 1080p. Sora has Remix, Re-cut, Blend and Storyboard tools but no start or end frame, no motion strength slider, no camera control and no region masks. No audio generation. Access is through Sora.com for ChatGPT Plus or Pro subscribers.

Luma Dream Machine (Luma AI, documented 2025-12-18) accepts JPEG or PNG up to 10 MB. Native clip length is 5 seconds at 720p. It supports start frame, optional end frame, and camera control presets (pan, tilt, zoom, orbit). No motion strength slider, no region masks, no audio generation.

Pika 2.2 (Pika Labs, documented 2025-12-18) accepts JPEG or PNG up to 10 MB. Native clip length is 4, 6 or 8 seconds at up to 1080p. It has motion strength (low/medium/high), camera control presets, region masks via Pikaffects and Scene Ingredients, and start frame image input. Pika 2.2 generates sound effects alongside the video but not speech.

Seedance 1.0 (ByteDance, documented 2025-12-18) accepts JPEG or PNG up to 15 MB. Native clip length is 2, 4 or 6 seconds at up to 1080p. It has a motion strength slider (0-100), camera control presets, and start frame. No end frame, no region masks, no audio generation.

Hailuo AI (MiniMax, documented 2025-12-18) accepts JPEG or PNG up to 20 MB. Native clip length is 6 seconds at up to 1080p. It has motion strength (0-10), camera control presets, and start frame. No end frame, no region masks, no audio generation.

Wan 2.2 (Alibaba, documented 2025-12-18) accepts JPEG or PNG up to 10 MB. Native clip length is 2, 4 or 8 seconds at up to 1080p. It has motion strength (0-100), camera control presets, and start frame. No end frame, no region masks, no audio generation.

LTX-Video 0.9.5 (Lightricks, documented 2025-12-18) is an open-weights model that accepts JPEG or PNG at up to 768×512 or 512×768. Native clip length is configurable from 2 to 6 seconds. It supports start and end frame input but has no motion strength slider, no camera control and no region masks. Apache 2.0 licence allows commercial use. Minimum VRAM is 12 GB for 768×512 at 121 frames.

Stable Video Diffusion 1.1 (Stability AI, documented 2025-12-18) is an open-weights model that accepts JPEG or PNG at 1024×576. Native clip length is 2 to 4 seconds. It has motion bucket ID (1-255) to control motion amount and configurable FPS. No end frame, no camera control, no region masks. The licence is non-commercial research use only.

Model	Max Source Resolution	Max Native Clip Length	Audio Generation	Start Frame	End Frame	Region Mask	Camera Control
Veo 3	4K	8 s	Yes	No	No	No	No
Kling 2.5	10 MB file	10 s	No	Yes	Yes	Yes (motion brush)	Yes (presets)
Runway Gen-4	4K	10 s	No	Yes	No	No	Yes (pan, tilt, zoom, dolly, crane, roll)
Sora	20 MB file	12 s	No	No	No	No	No
Luma Dream Machine	10 MB file	5 s	No	Yes	Yes	No	Yes (presets)
Pika 2.2	10 MB file	8 s	Yes (sound effects)	Yes	No	Yes (Pikaffects)	Yes (presets)
Seedance 1.0	15 MB file	6 s	No	Yes	No	No	Yes (presets)
Hailuo AI	20 MB file	6 s	No	Yes	No	No	Yes (presets)
Wan 2.2	10 MB file	8 s	No	Yes	No	No	Yes (presets)
LTX-Video 0.9.5	768×512	~6 s	No	Yes	Yes	No	No
Stable Video Diffusion 1.1	1024×576	~4 s	No	Yes	No	No	No

The uncanny valley for still-portrait animation is not a single problem. It is four distinct failure modes that appear at different points in the generated clip. Know each one before you generate.

Identity Shift

This is the most alarming failure. The person on screen starts as a recognisable version of the source image and gradually changes ethnicity, age or gender over the clip duration. The model interpolates toward the statistical distribution of faces in its training data rather than anchoring to the source. The US Copyright Office (report published 2025-01-29) has noted that selection and arrangement of AI outputs may be copyrightable, but a generated face that shifts identity has no stable authorship. The fix is shorter clips. Kling 2.5 offers 5-second generation. Runway Gen-4 offers 5-second. Sora offers 4-second. Use the shortest available native clip length. Identity shift becomes visible after roughly 4 seconds on most models.

Uncanny Mouth and Eye Movement

The mouth opens and closes but the timing does not match the words described in the prompt. The eyes blink at random intervals or do not blink at all. The skin around the mouth stretches in a way that no human face does. This is caused by the model having no articulatory model of speech. It has learned pixel patterns that correlate to motion but has no phonological understanding. Pika 2.2's region masks let you constrain the animation to the lower half of the face, which reduces the chance that the eyes drift. Veo 3 generates audio alongside the video but its documentation (2025-12-18) states the motion is controlled by the text prompt alone. You cannot feed an audio file to correct the lip sync timing.

Limb Mutation

If the frame includes shoulders, the shoulder-to-neck boundary is a site where arms warp, multiply or detach. The reason is structural: the model has no skeletal model of the subject. It predicts pixel displacement from a single reference frame. The shoulder edge is a high-contrast boundary where the model has the most freedom to hallucinate motion. Crop the source image below the collar. This removes the boundary entirely.

Flicker and Temporal Inconsistency

Frame-to-frame brightness shifts, skin texture that softens and hardens every few frames, and background colour changes that have nothing to do with the intended motion. This is the model's latent-space decoder producing inconsistent results across frames. The claimed resolution of a model is often not the effective resolution. Independent frame analysis shows that a 1080p output from many models contains effective detail closer to 720p upscaled. The EU AI Act transparency obligations (Article 50(2) and 50(4), applicable 2026-08-02) require deployers to label synthetic video. A flickering output will be detected by platform classifiers before a human viewer notices it.

Animating a still portrait of a real person without that person's explicit consent is a legal exposure that every model's documentation addresses differently.

Veo 3 (documented 2025-12-18) prohibits generation of photorealistic identifiable individuals without approval. Runway Gen-4 prohibits generation of public figures without consent and deceptive deepfakes. Sora prohibits uploads of people without consent and political deepfakes, and embeds C2PA metadata on all outputs. Luma Dream Machine prohibits photorealistic depictions of identifiable minors and non-consensual intimate imagery. Pika 2.2 prohibits generation of public figures for deceptive purposes. Seedance 1.0 prohibits public figure deepfakes and requires real-name verification in China. Wan 2.2 requires real-name verification in China. Kling 2.5 prohibits political figures. Hailuo AI prohibits political figures and deceptive deepfakes. LTX-Video 0.9.5 and Stable Video Diffusion 1.1 have no built-in safety filters in their open weights, placing the entire consent obligation on the user.

The YouTube Creator Studio labels policy (effective 2024-03-18) requires disclosure when realistic content is made with altered or synthetic media. The label reads 'Altered or synthetic content' and appears in the video description panel. If the subject of the generated video does not know their likeness is being used, the uploader faces both platform takedown and legal liability.

Newcomers assume the output will be physically consistent. The model does not understand that a face has two eyes, a nose and a mouth that stay in fixed relation to each other. It predicts pixel motion from patterns learned during training. The result is that the most carefully framed source image can produce an output where the person on screen is unrecognisable by the end of the clip. The only reliable workaround is to generate the shortest possible clip and accept that the output will not match the source for the full duration. Plan for a 4-second usable clip from a model that advertises 10-second generation. That gap between claimed maximum clip length and usable clip length is the defining constraint of still-portrait to talking-head work today.