Imagen, Veo, Runway, Sora: a director's guide to choosing models in 2026
Each generative model has a personality. The personality matches some films and fights others. Here's how to pick.
There is no best model. There are six or seven good ones, each with a personality, and the question is which one's personality matches the film you're making. This is what cinematographers used to argue about with film stocks. The argument is back, just louder and faster.
Imagen — Google's still
Imagen's strength is photographic restraint. It renders skin, textiles, and naturalistic light without the shimmer that gives most generative stills away. It struggles with hands, with text, and with stylised illustration — anything that asks for departure from photoreal. Use Imagen for keyframes when you want the storyboard to read as photographed reference. Don't use it for paintings, drawings, or anything that should look composed by a human hand.
Best paired with the Documentary, Intimate, and Melancholic tone presets. Worst paired with Dreamlike — the photoreal floor undercuts the surrender to logic that dreamlike depends on.
Veo — Google's video
Veo 3 is the first generative video model that handles long takes (10s+) without the subject morphing mid-shot. It's especially strong on landscape, vehicle interiors, and slow-motion. Weak on dialogue scenes because lip-sync is still synthetic. Strong on cuts that lean on light and movement.
Best paired with Epic and Tense tones — the model's instinct toward cinematic motion suits both. Worst paired with Documentary, where Veo's polish reads as performed.
Runway — Gen-3 / Gen-4
Runway is the painter of the bunch. Highest stylisation, most reliable on art-directed footage, easy to push into specific looks (anamorphic flare, film grain, 4:3 academy crop). Lower realism than Veo but a wider expressive range. Strong for music videos, dream sequences, anything where artificiality is the point.
Best with Dreamlike, Playful, and Pastel-style films. Pairs poorly with Documentary.
Sora — OpenAI's video
Sora's signature is camera awareness — it knows what a dolly does, what a Steadicam looks like, when to push in. The motion vocabulary is the most consciously cinematic. Cost-per-second is highest of the bunch.
Use Sora for key beats where the camera move IS the shot. Don't waste budget on static-frame fillers; cheaper models do those just as well.
Pika & Seedance
Pika and Seedance occupy the speed lane. Pika is the most consistent on character motion (hands, gestures, body mechanics). Seedance is the strongest on image-to-video — feed it a still and it animates the world around the subject without redrawing the subject.
Use them as the workhorse models for ten-second blocks where you don't need the prestige of Sora or the polish of Veo. The cost saving is real.
On the meta-question
Studio Pro routes through whichever model fits the modality and tone — you don't pick per shot unless you want to. The advantage of the orchestrator is that it picks based on cost-quality fit you'd otherwise have to relitigate per scene. The advantage of BYO keys is that you keep the relationship — and the bill — direct.
Either path is correct. The wrong path is paying full Sora prices for a static frame.