The craft: consistency and control

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. is the problem the field spent years failing to solve, and the current everyday answer is reference-image conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image.. Runway's Gen-4 (March 2025) lets you supply reference images of a character or scene and generate consistent versions across new shots, angles, and lighting with no per-character training, and Google's Veo 3.1 (October 2025) added the same idea under the name reference images, or ingredients, accepting a few references of a character, object, or scene to steer generation. The workflow is now: lock identity in one strong reference, then reuse it for every shot. When a reference is not enough for a lead who must survive dozens of scenes, pros still train an identity 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., a small model fine-tuned on one subject. One published method (from the AI-film school Curious Refuge) is concrete: collect at least ten varied images of the subject, avoid multiple frames from the same shoot, keep faces unobscured, use the character's name as the trigger wordTrigger wordA unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style., and train roughly a thousand steps, a job of minutes on a hosted service. Treat those numbers as one credible recipe, not a universal standard.

Motion and camera are directed two ways, through the driving image and through language. Camera-control prompting, the directorial vocabulary of dollies, cranes, whip pans, push-ins, dolly zooms, and rack focus, is now understood well enough that tools market those moves as callable from the promptPromptThe text description you provide to a model to specify what you want it to generate.. This is where a real cinematography vocabulary pays off literally: the knowledge that lets a director say slow push-in on a fifty millimeter, shallow depth of field is the same knowledge that steers the model, which is why directing language transfers almost verbatim. First-and-last-frameFirst-and-last-frameA video workflow where you fix the opening and closing frame and the model generates only the motion between them. keyframing is the strongest form of control: you provide the opening and closing image and the model synthesizes the transition, turning an unpredictable generation into a bounded interpolation that starts and ends where the edit needs it. Shot extensionShot extensionGenerating a new clip that connects to the end of a previous one, a way to build continuous motion past a model's short length ceiling. addresses the length ceiling by generating clips that connect to a previous one, so continuity is engineered clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. by clip rather than captured in one take.

Performance is now its own craft stage. Runway's Act-One (October 2024) drives a generated character from a single actor's video and voice, preserving eye-lines, micro-expressions, pacing, and delivery, from a simple single-camera setup with no motion-capture rig or manual face-rigging, and Act-Two (July 2025) extended capture to full head, body, and hand gestures from a driving video plus a reference character. This is modern puppeteering: the actor's job survives, the mocap stage and the rig do not. It is also why the strongest AI films route their most demanding moments through a real human performance rather than trying to synthesize emotion from text.