Generative AI: The 2026 Field Guide

The second major frontier is character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work., the problem of generating multiple images of the same character that all clearly show the same person, with the same facial features, body type, and visual identity, across different poses, settings, and styles. This is the single hardest unsolved problem in generative imagery for narrative use cases (comics, storyboards, animation pre-production, longform video) and the technique that solves it well will unlock an entire category of applications that do not really work yet.

ANALOGY

Imagine a police sketch artist who has never met your friend. You describe the friend in words, the artist draws a sketch, and the result is recognizably a person with the right hair color and rough features but it is not really your friend, you would not pick that face out of a crowd. Now imagine you give the sketch artist one photograph of your friend before describing them. The next sketch is much closer. Now imagine you give the artist forty photographs and let them practice drawing your friend a hundred times. Now the sketches start to look like your friend in every pose. Character consistency is the gap between those three scenarios. IP-Adapter is the one-photo case, decent but loose. Subject LoRAs are the forty-photo case, much tighter but expensive to set up per character. The frontier research is trying to get the one-photo case as good as the forty-photo case, which would unlock comics, storyboards, and AI-generated narrative video as practical applications for the first time.

The current solutions are partial. Intellectual Property (IP)-Adapter, introduced in 2023 by Tencent's research team, lets you provide a reference imageReference imageAn image fed to a model as conditioning to carry a subject, style, or composition into new output, rather than describing it in words. alongside your text promptPromptThe text description you provide to a model to specify what you want it to generate. and biases the generation to look like the reference. It works well for capturing general identity (this is the same person) but struggles with fine details (it is not exactly the same person). Subject LoRAs, covered in the Low-RankRank (LoRA rank)The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files. Adaptation (LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models.) section, are the other main approach, train a custom LoRA on 20-40 photos of the subject and use it for all generations. This works much better but requires per-character training and produces drift across generations within a single project.

The frontier work in 2025-2026 has produced more sophisticated approaches. StoryDiffusion, introduced by a team at Nankai University and ByteDance, reformulates the diffusion process to maintain consistency across a sequence of images by sharing attention between them, instead of generating each image independently, the model jointly attends to all images in the sequence and enforces feature consistency. The result is that a generated 'comic strip' has visibly the same character across panels without requiring per-panel training. Concurrent work like InstantIDInstantIDA technique for preserving a specific person's identity across generations from a single reference photo, without training a LoRA. and PhotoMaker pushes the IP-AdapterIP-Adapter (Image Prompt Adapter)A method for conditioning generation on a reference image, so a subject or style from a picture carries into new outputs. approach further, getting closer to one-shot identity preservation from a single reference image.

The other thread is from the video model side. Sora's technical report claimed emergent character consistency over short clips. Veo 3 demonstrably handles character consistency over its 8-second clips. The mechanism is different from the image-side techniques, for video, the model is naturally producing a temporally coherent sequence and identity preservation is a side effect of the underlying video architecture. The interesting research question is whether the techniques that work for video can be back-ported to images. Early results suggest yes, and the next generation of image models (FLUX 2, presumably; SD4, presumably) will probably have meaningful character consistency built into the base model rather than requiring per-character LoRAs.

What changes when this is solved

Once character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work. works reliably, several new product categories become viable. Animated narratives produced by individual creators become possible at scale (right now they require either heroic per-frame curation or an animation studio). Character-driven advertising campaigns can be produced without hiring the actors who would normally be the consistency anchor. Comic and graphic novel production becomes faster by orders of magnitude. AI-generated TV shows with persistent characters become a real possibility. The 'visual narrative' space is currently bottlenecked by character drift, and removing that bottleneck is going to release a wave of products.

Character consistency and identity preservation

What changes when this is solved

Get the week's moves in your inbox.