AI video's benchmark shifts to holding a character across cuts

The benchmark for AI video moved this year. A model is no longer judged on one impressive shot. It is judged on whether it can hold a character, a voice and a look across a sequence of cuts, and the tools have changed shape to match. Reference images, saved characters, storyboards, keyframes and performance footage have become standard inputs, where a fresh paragraph of text for every shot used to be enough.

The change shows up in the makers' own documentation. Kling's VIDEO 3.0 model guide lists Multi-Shot and Start Frame plus Element Reference as capabilities the earlier 2.6 model did not have. Higgsfield, a vendor that sells character tools of its own, describes the underlying problem in its guide to consistency tools: AI video models generate every shot independently, with no memory of the one before, so text descriptions reduce drift but do not stop it. Only an identity anchor does, the guide says.

Why consistency replaced the pretty clip

Within a single generation, consistency is not the hard part. As Higgsfield's guide explains, a diffusion model generates all frames of a clip in one pass, treating the whole sequence as one context, so a face, its lighting and its geometry stay stable through camera moves and angle changes. The problem starts at the next generation, which is a fresh context with no carry-over. The same prompt twice produces two people who look almost, but not quite, alike.

That is why the tools below now sell an anchor rather than a prompt. An anchor is something concrete the model can hold onto between generations: a photo, a saved identity, a storyboard frame, a performance clip. Each maker built a different kind of anchor, and their own documentation shows where the approach still fails.

What each maker actually built

Kuaishou's Kling 3.0 builds a reusable character from a few images. Kling's Element Library guide says an Element is a composite asset built from at least 2 and up to 4 reference images, one main front-facing image plus supplementary angles. A character Element can instead come from a 3 to 8 second video clip of a single person, from which the model extracts the appearance and the native voice. For character Elements, Kling lets you bind a voice, either from an uploaded audio clip or from its existing Voice Library, so the same face and the same voice travel together across projects. The same guide says video generation supports up to 7 reference characters and image generation up to 10, and that VIDEO 3.0 can bind up to 3 elements when you supply a start frame or start and end frames. The model guide puts the output at up to 15 seconds per generation.

Runway anchors a character from a single image. Runway's Gen-4 References guide says the feature "excels at generating consistent characters across different lighting conditions, locations, and treatments, all from just a single reference image," and that you can select up to three active references for a single generation. Its video model, Gen-4.5, currently offers Text to Video and Image to Video control, with more inputs "coming soon," according to the Gen-4.5 help page. Runway's research page says Gen-4.5 ranked first on the Artificial Analysis Text-to-Video leaderboard with 1,247 Elo points, ahead of Veo 3 at 1,226 and Kling 2.5 at 1,225, on a leaderboard dated November 30, 2025. For editing, Runway's Aleph 2.0 is an in-context video model: you edit one frame, and it modifies the rest of the video to match, changing only what you asked for while preserving the background and lighting.

Google's Veo 3.1 interpolates between two frames. Google's Gemini API documentation says Veo 3.1 lets you create videos by specifying the first and last frames, filling in the motion between two images instead of inventing the whole shot. The same page says Veo 3.1 accepts up to three reference images to guide the content of a generated video.

Luma's Ray3 keeps or replaces a character in existing footage. Luma's Ray3 Modify guide describes a set of video-to-video and Character Reference features. With an input video plus a character reference, Ray3 replaces the character in the footage with your reference and leaves everything else unchanged, adapting the reference to the original video's lighting and style. A keyframe and character reference combination is described as the highest level of control, letting you set the start or end frame while the character reference holds the identity.

Vidu's newest model takes the most references. Shengshu Technology's launch announcement for Vidu Q4 Preview says the model supports up to 15 image references and up to 3 audio references, with the image references meant to define characters, wardrobe, props, products and environments in one setup. SLOP TV covered that launch separately; it is worth noting here only because the reference count is the largest of the group.

Where it still breaks

Two characters in one shot remain the hard case. Higgsfield's guide states it directly: multi-character interaction is unsolved across all platforms. Two characters hold their identities in isolation, the guide says, but when they share a close-up or physically interact, identity blurring appears at the intersection points. Its FAQ answers the same question flatly: two characters interacting in the same shot still produce identity blurring on every platform as of mid-2026.

There is a real tension between that claim and the makers' marketing. Kling's own model guide lists "Multi-Character Coreference (3+)" as a capability new to VIDEO 3.0, and its Element Library guide says that in complex multi-character scenes the model can lock and maintain each character's features. The vendor is claiming progress on the exact case a competing vendor says is unsolved. Both are claims by parties with products to sell, and neither has published a test a reader can check.

Long sequences still drift. The same Higgsfield guide says long-form consistency degrades past 30 seconds on most models, with expression repetition and subtle drift appearing even on stronger outputs, and that dynamic action sequences need multiple regeneration attempts on every tool it lists. A character who holds for a handful of short shots is routine now. A character who holds across a full narrative, built from many clips, is not.

What to do about it this week

Pick one anchor and reuse it. The consistent advice across the makers is to stop describing the character in words and start feeding the model a fixed image on every clip, with the prompt controlling only what the character does and where. Kling's guide recommends a front-facing reference image for higher consistency, and clean background audio with a neutral voice for voice binding.

Group your shots by type. Keep a character's close-ups together, then the wider shots, so the model re-establishes the identity less often across the sequence. Use keyframes when a shot needs a specific start or end, the way Veo 3.1 and Luma's Ray3 both allow, and use a storyboard or multi-shot mode when you want the cuts planned for you.

When a shot genuinely needs two characters, generate them separately against their own references and composite them, rather than asking one generation to hold both. That workaround is the practical consequence of the blurring problem above, and it is the state of the tooling until a vendor publishes a test showing otherwise.

Try one workflow this week: build a single character Element in Kling from a front-facing photo and a clean voice clip, then generate three short shots with it and check the face and voice against the reference before you generate anything longer.