Character.ai post-trains its own CAI-Image models on Qwen-Image

Background

Character.ai describes its platform as built on fandom: people bring characters to life there, and those characters are now being carried into comics and, soon, into video ^[1]. The company says millions of people create characters, scenes and stories on the platform every day, and that most of this creation happens in a single tap ^[1]. That kind of use places a specific demand on image generation. A story asks for the same character again and again — in a different art style, from a different angle, alongside another character, on the next page — and a reader notices immediately when the face or the outfit drifts.

What was announced

On 16 September 2026 the company said it had built its own image models instead of relying on general-purpose ones ^[1]. The family is called CAI-Image, and its members are post-trained versions of the open-source Qwen-Image models, tuned for character-driven storytelling ^[1]. Post-training here means starting from a base model that has already learned to generate images in general, then continuing to train it on a narrower set of targets: the base supplies general image competence, and the extra training bends it toward the job at hand.

What the models were tuned for

The stated goal is that the same character stays recognisable through every style, scene and page a story asks for ^[1]. The team lists four areas of work: style range and transfer, multi-character scenes and pose control, cinematic camera control, and manga and comics generation ^[1]. Style range and transfer is about moving a character between visual treatments without losing identity; multi-character scenes and pose control are about placing more than one character in a frame and choosing how they stand; cinematic camera control is about framing a shot; manga and comics generation is about the panel layouts and line work of comic pages.

Where it is used

CAI-Image draws the recently launched (c.ai) Comics and powers image generation across surfaces such as Imagine and Imagine Message ^[1]. It is therefore not a standalone demo: it sits underneath features people already use to make characters and stories, and the comic-making feature is built directly on top of it.

Why the company says it matters

Character.ai argues that this approach delivers generation in seconds, a fraction of the cost, and control from training data to serving ^[1]. The first two claims are about serving economics: a model built in-house and tuned to a narrow job does not have to hand every request to a general-purpose model. The third is about scope: because the team works on both the training data and the serving path, changes to the model can be made within one stack.

What it signals

Choosing to post-train rather than rely on general-purpose models puts the character-consistency requirement inside the model itself. The same stack draws comics today and, according to the company, is meant to carry characters into video next ^[1]. For a platform whose centre is characters and fandom, owning that capability means the path from chat to comics and then to video can run along a model line the company controls ^[1].