Pick a video model from a year ago, ask for the same character in two different shots, and watch their face change between cut one and cut two. That was the entire problem with AI filmmaking. It still is, technically. The models still drift. What changed is that the people making real shorts have quietly stopped relying on the model to remember who anyone is.
Character consistency is the difference between a clip and a movie. A clip can be one good face with motion. A movie needs that face to walk into the next room and still be the same person. Until late last year, the practical state of the art was regenerate, pray, re-roll, accept the closest match. That is not a workflow. It is gambling. What changed is not the models. It is what creators do around them.
The keyframe-first pipeline took over
The pattern that took over among working AI filmmakers in the past six months is image-to-video, anchored on a deliberate keyframe. You do not start with a text prompt to a video model. You start with a single high-fidelity portrait, usually generated in Flux 2 or Midjourney from a tight character bible, and treat that image as the canonical truth of who the character is. Every subsequent clip generates from that frame, not from a description.
The shift sounds small. It is not. Text-to-video burns the entire identity question into a guess on every generation. Image-to-video burns it once and then asks the model to add motion. Drift drops by an order of magnitude. The trade-off is that you lose some compositional surprise; the model cannot reinvent the framing if the framing is already locked. Working filmmakers consider that a feature.
Anchoring across more than one clip
A single keyframe holds for one clip. Holding across forty clips is harder. Two practical answers have settled in.
The first is reference banks. Instead of one portrait you build five: front, three-quarter, profile, lit warm, lit cool. You feed the right reference into each shot. This catches most of the rotation and lighting drift that breaks short sequences.
The second is what Kling shipped last week in 3.0 as "Bind Subject". It works differently from a reference image: rather than treating the photo as a 2D pattern to imitate, the model maps the subject as a 3D anchor and reasons about how the same jacket should fold when the person turns. It is the first feature in this category that holds up across a multi-shot scene without per-clip babysitting. Runway has a comparable mode in Gen-4.5. Pika's character pass is meaningfully better than its reputation suggests. Higgsfield and OpenArt sell entire products built on this idea.
Where the workflow still breaks
Three failure modes are still real, and pretending otherwise wastes everyone's time.
Fast motion warps faces. A character turning their head 180 degrees mid-clip will lose continuity at the rotation point on every consumer model. The fix is to stop asking for that motion. Cut on the turn instead.
Long clips drift. Even with bind-subject locked on, generations beyond eight to ten seconds accumulate small errors that look fine in isolation and obvious next to the previous shot. The fix is to chain three-to-four-second clips with deliberate keyframes, not to push the engine for length.
Wardrobe and prop changes confuse the anchor. A character takes off a jacket in one shot and the jacket comes back in the next. There is no clean fix yet. The workaround is treating wardrobe state as part of the keyframe and never expecting the model to track it across cuts.
What to do this week
If you are starting a sequence, the practical recipe is straightforward. Lock a character bible on paper first. Generate a five-image reference set at high resolution. Build every shot from the closest-matching reference. Keep individual generations under five seconds. Cut whenever the character has to do something physically complex. If your tool does not expose start-frame, end-frame, and a subject-binding mode, use one that does. Running this workflow without those controls means you are back to praying.
That last point is the one most worth hearing. Character consistency in 2026 is not a model problem. It is a pipeline problem. The labs will keep improving and every quarter the drift gets a little smaller, but the gap between "shipped a clip" and "shipped a scene" closes only when the workflow does the heavy lifting the model cannot. Promvie's pipeline is built on exactly that assumption: the keyframe, the reference set, and the model assignment per shot are the product, not afterthoughts.
The interesting question this year is not which model is the most consistent. It is which pipeline you have built around them. Pick one that lets you anchor.