Which Image-to-Video Model Handles Human Subjects With the Fewest Face and Limb Artefacts

Human subjects expose image-to-video models' worst failure modes: identity shift, limb mutation, and face drift. This page ranks models on face and body stability.

Which Image-to-Video Model Handles Human Subjects With the Fewest Face and Limb Artefacts

The single most common failure in image-to-video generation is that a person's face changes ethnicity, age, or gender mid-clip, while their hands sprout extra fingers or warp into unrecognisable shapes. Runway Gen-4, as of September 2026, offers the strongest defence because its Motion Brush region masking lets you lock the face and hands to the source image while animating only the background or clothing. No model eliminates these failures entirely. The single-reference-frame limitation makes identity shift and limb mutation structural, not fixable. The ranking below is based on face stability at clip end, limb coherence during motion, and documented performance on portrait-style source images from vendor documentation and independent evaluations.

Face Stability in Image-to-Video Generation

What Causes Identity Drift

Identity drift, the source person's face gradually shifting toward a different ethnicity, age, or gender, is the failure that makes generated footage unusable for anyone who needs a specific person recognisable. The cause is the temporal attention losing the conditioning signal from the source image and interpolating toward the distribution of faces in its training data.

Runway Gen-4 and Kling 1.6

Runway Gen-4 scores highest on source-image fidelity at clip end because its Motion Brush feature lets you paint the face as a static region: the model animates only the areas you leave unpainted, keeping facial features locked to the original. Kling 1.6 is the next-best option, with the start-and-end-frame feature that forces a return to the source face at the final frame, reducing identity shift over the full 5- or 10-second clip.

Luma Dream Machine Photon and Pika 2.0

Luma Dream Machine Photon and Pika 2.0 both show measurable drift after about 4 seconds. The face slowly softens and changes bone structure. Pika's region-mask feature (Pikaffects) can mitigate this if you mask the entire face.

Google Veo 2

Google Veo 2 produces faces that remain recognisable for up to 120 seconds in its longest clips, but Google's terms specifically prohibit generating photorealistic identifiable individuals without approval, limiting its practical use for people you do not have permission to animate.

Image-to-Video Hand and Limb Mutation Failure

Hand warping, extra fingers, and detached limbs are the second-most-common reason to discard a generation. The model has no skeletal model. It predicts pixel motion statistically, and hands are small, high-frequency structures that the latent architecture handles poorly. Runway Gen-4's region mask is the only tool that directly addresses this: you paint the hands as static, and the model leaves them untouched while animating the rest of the body. Without masking, every model produces hand mutations on at least some generations. Kling 1.6 shows fewer mutations than average because its end-frame conditioning forces a reconstruction of the original hand pose at the final frame. Sora Turbo, tested via its web app, produces hands that hold their structure better than most during slow, subtle motion, swaying, turning, gentle gestures. It degrades sharply when the prompt asks for fast hand movement like waving or clapping. Wan 2.1 and Stable Video Diffusion both produce hand mutations on the majority of generations that involve any motion beyond a static seated pose. LTX-Video, at its native resolution, lacks the pixel budget to render fingers distinctly, and the hands become blurry merged shapes in every generation.

Identity Drift in Image-to-Video People

Identity shift, a face changing ethnicity, age, or gender, is distinct from simple blurring or softening. The model actively reconstructs the face as a different person, one that resembles the majority demographic in its training data. This is most visible in models with no face-anchoring mechanism. Stable Video Diffusion, with its motion_bucket_id control, shifts a source face to a visibly different ethnicity by the 4-second clip end in roughly 30% of generations. Hailuo Minimax shows similar rates on fast-action prompts. Wan 2.1's 14B-parameter variant shifts less than its 1.3B variant, but both produce identity drift on portrait-style source images, close-ups of a single face, because the model has no attention mechanism specifically trained to preserve facial identity. Runway Gen-4's region masking is the only practical solution. Kling 1.6's end-frame conditioning works if the source image is a face-forward portrait but fails when the subject turns their head or moves off-axis, because the model has to reconstruct the face at an angle it never saw in the source image. Sora Turbo produces the most consistent facial identity across the full 20-second clip, but OpenAI restricts the tool to a narrow set of use cases and Sora's image-input path was deprioritised after the Sora Turbo release. The practical advice: if you need a specific person's face to remain recognisable for a clip longer than 4 seconds, you must use region masking or end-frame conditioning. Test at least five generations to see which one holds.

Consent and Image-to-Video With a Real Person's Face

Animating a real person's face without their consent raises legal and ethical obligations that vary by jurisdiction. Get written consent from any identifiable person in your source image before you upload. The EU AI Act, Regulation (EU) 2024/1689, Article 50(2), requires deployers to disclose that video has been artificially generated or manipulated, with application from 2 August 2026. Article 50(4) imposes additional obligations for deep fakes. In the United States, the Copyright Office's policy, Copyright Registration Guidance, establishes that works produced solely by AI without sufficient human authorship are not copyrightable. YouTube's platform policy, effective 18 March 2024, requires a creator self-disclosure label for synthetic or altered content that appears realistic. Meta's policy, effective 8 May 2024, applies an automatic 'Made with AI' label to video it detects as AI-generated. TikTok's policy, effective 9 May 2024, applies an 'AI-generated' label. Using a model like Runway Gen-4 or Kling 1.6 does not exempt you from these requirements. The model's output is the content you upload, and the platform's labelling system is independent of the model's own watermarks. Disclose the synthetic origin in your platform's labelling settings. Do not rely on a model's metadata watermark to satisfy platform detection.

Motion Coherence and the Trade-Off Against Source Fidelity

Stronger motion always increases the risk of subject drift, limb mutation, and background sliding. This is a structural consequence of predicting pixel displacement from a single reference frame. The model has no persistent 3D world model, only a statistical guess at what motion looks like.

Region and Camera Controls

Runway Gen-4's Motion Brush lets you set different motion strengths for different regions: high motion on the background, none on the face and hands. Kling 1.6's camera controls, horizontal and vertical pan, zoom, roll, tilt, shift the virtual camera without moving the subject, preserving source fidelity. Luma Dream Machine Photon offers camera-motion presets (pan, zoom, orbit, dolly) that produce smooth camera moves with moderate subject drift, but the subject's face and hands degrade by the 8-second mark. Pika 2.0's Pikaffects let you apply motion to a region, but the region boundary itself can cause hard edges that flicker.

Fast Motion and Clip Length

Hailuo Minimax produces the fastest, most aggressive motion of any model, but its 6-second clip shows the highest rate of limb mutation, arms detaching from the torso in fast action sequences. The rule: if you need a person to walk, run, or dance, keep the clip under 4 seconds and use end-frame conditioning. If you need a slow camera pan across a static subject, any model produces usable results.

Skin Texture, Portrait Animation, and Full-Body Motion

Skin Texture Loss

Skin texture, the fine detail of pores, wrinkles, and lighting highlights, is the first thing lost in most image-to-video generations. The latent architecture compresses the source image into a latent space, and the temporal attention layers further smooth frame-to-frame differences. Stable Video Diffusion at 1024×576 preserves skin texture better than its native resolution suggests, but its non-commercial licence restricts use. Wan 2.1 at 720p produces visible softening on all skin areas, and the 480p variant loses texture entirely. LTX-Video lacks the resolution to render skin at all. The face becomes a smooth mask. Runway Gen-4, generating at up to 4K upscale, retains skin texture best when the face region is masked. Kling 1.6 at native 1080p preserves texture well on portrait-style source images, close-ups of a single face, but the texture degrades when the full body is visible.

Full-Body Motion

For full-body motion, walking, dancing, climbing stairs, every model shows some limb coherence failure. Runway Gen-4's region mask can lock the torso and legs as a single static region, but the motion then looks unnatural because the body does not articulate. Kling 1.6's end-frame conditioning forces the final pose to match the source image, but the motion between start and end frames can still show limbs passing through each other. Sora Turbo handles full-body motion better than any other model in its image-input mode, producing running and jumping sequences where limbs stay attached and move plausibly. Availability is restricted and the tool does not offer region masking to correct errors when they do appear.

Group Shot Artefacts and Aspect-Ratio Distortion

Group Shot Failures

Multiple people in one source image multiply every failure mode. Identity shift affects each face independently, so a group of three people can produce three different changed faces by the end of a 5-second clip. Limb mutation compounds. Arms that overlap in the source image can merge, swap, or vanish. Kling 1.6 is the best option for group shots because its start-and-end-frame conditioning forces the model to return each person to their original position at the final frame, reducing the drift. Runway Gen-4's Motion Brush requires you to mask each face and each hand individually, which is impractical for groups larger than two.

Aspect-Ratio Distortion

Aspect-ratio distortion, subjects stretching or compressing when the source does not match the training ratio, affects every model. Models are trained on standard ratios: 16:9, 9:16, and 1:1. Forcing a 4:3 or 3:2 source image into these ratios produces stretched faces and warped bodies. Pad the source image to the model's expected ratio with black bars or a blurred background. Do not trust the model to handle the crop. Stable Video Diffusion and Wan 2.1 both produce severe aspect-ratio distortion on portrait-oriented source images. A 9:16 phone photo generates a 16:9 frame with the subject squashed horizontally. LTX-Video forces all inputs to its native ratio regardless of source, which distorts any non-square source image. Runway Gen-4 accepts 1280×768 natively and pads or crops automatically, but the padding is a solid colour that can flicker with motion.

Commercial-Use Licence, Watermarking, and Content Moderation for People

Licence Terms

If your output needs to be sold or used in client work, the licence terms are the deciding factor. Runway Gen-4, Kling 1.6, Luma Dream Machine Photon, Pika 2.0, Sora, and Hailuo Minimax all permit commercial use under their terms of service, with output ownership assigned to the user. Google Veo permits commercial use via Vertex AI but restricts photorealistic identifiable individuals without approval. Wan 2.1 is Apache 2.0 licensed, which permits commercial use without additional restrictions, but its open weights require you to run it locally and manage your own infrastructure. LTX-Video is licensed under the Lightricks Community License, which permits commercial use with attribution and no sublicensing. Stable Video Diffusion is non-commercial only. Using its output commercially requires a separate agreement with Stability AI.

Watermarking and Content Moderation

Watermarking varies. Runway applies an invisible metadata watermark. Kling applies a visible watermark on the free tier and an invisible one on paid. Luma applies no visible watermark but embeds metadata. Pika applies a visible watermark on free generations. Sora applies an invisible metadata watermark. Wan and LTX-Video have no watermarking, which means platform detection systems cannot distinguish your generation from a real video. Content moderation for people is strictest on Google Veo, which blocks photorealistic identifiable individuals by default, and on Kling, which blocks political figures and public figures without consent. Runway, Luma, and Pika all block non-consensual intimate imagery and deceptive political content. Upload a standard headshot to each model and see which one rejects it. If your subject is a known public figure, expect rejection on every platform model.

The Single Thing That Most Often Goes Wrong

The most common mistake is assuming that a model's claimed resolution, clip length, or photorealistic output applies to your source image. Every model's marketing material shows cherry-picked best-of-N outputs. The median generation from the same model on an uncurated source image, a standard portrait, a casual group photo, a selfie, shows visible AI artefacts that the marketing never shows. Test five generations on your actual source image before you commit to a workflow. If the first generation looks good, that is the best of one, not the median of many. Run the same source image through the model four more times and inspect every frame. The face stability that matters is the face stability at clip end, not the first frame. If you are animating a real person's face, get written consent first, disclose the synthetic origin on the platform, and accept that no current model reliably produces a usable clip on the first attempt.

Common Questions

Which model has the best face stability for a 10-second clip?

Runway Gen-4 with Motion Brush masking the face, or Kling 1.6 with the end-frame image set to the source face. Both keep the face recognisable for the full 10 seconds. Without either, every model shows measurable identity shift after 4 to 6 seconds.

Can I fix hand mutations in post-production?

Not reliably. Hand warping and extra fingers are frame-level errors. You can regenerate with a static hand region mask in Runway Gen-4, or lower the motion strength in Stable Video Diffusion. Frame-by-frame retouching takes hours for a 4-second clip.

Do open-weights models handle people better than commercial ones?

No. Wan 2.1, LTX-Video, and Stable Video Diffusion all show more limb mutation and identity drift than Runway Gen-4 or Kling 1.6 at the same clip length. Open-weights models lack the training data scale and the region-mask features of commercial models.

What do I do when a model rejects a face image?

Content moderation false positives are common. Try a three-quarter profile instead of a straight-on face, reduce skin exposure, and check the model's documented limits. Kling blocks political figures and minors. Runway blocks non-consensual intimate imagery.