Which AI video model should I use on Martini?

If you have a source image and want it to move, you are doing image-to-video. The strongest models for this in 2026 are Seedance 2 (best motion fidelity to the source), Kling 2.5 (good prompt-following), and Luma Ray (fastest, cheapest). Drag your image into an Image node, connect it to a Video node, and pick one of those three from the model dropdown. The output preserves the look of the source image and adds motion driven by the prompt.

If you are starting from a text prompt with no source image, you are doing text-to-video. Most flagship video models support text-to-video, but the differences are larger here. Sora 2 and Veo 3 produce the most cinematic motion and best prompt adherence; Hailuo is the value pick for fast iteration. Start with one of the fast ones to nail the framing and prompt, then promote the final cut to a higher-cost flagship model.

Multi-shot video — a sequence of related clips telling a short narrative — is what Sora 2 and Kling 2.5 are designed for. They support per-shot prompting with consistent style across the sequence. On Martini, the multi-shot workflow lets you stage a row of video nodes that share style and character cues, and these two models give you the best across-shot continuity.

Character consistency — the same identifiable character appearing across multiple clips — is the harder problem. Vidu and Kling have the strongest cross-shot identity preservation when paired with a character reference image. Use the character-consistency feature on Martini, drop in the reference, and route through Vidu or Kling for a sequence where the character needs to be recognizably the same person.

Cinematic camera motion, depth of field, and physical realism are where Google Veo 3 and Sora 2 lead. If the brief is film-like or trailer-style, pay the higher credit cost — these models produce the kind of motion grammar that downstream editorial work depends on. For product, beauty, fashion, and ads, Seedance 2 and Runway Gen-4 produce clean commercial-style motion at lower cost.

Cost per second of video varies by an order of magnitude across models. Hailuo and Luma Ray sit at the cheap-and-fast end; Veo 3, Sora 2, and Runway Gen-4 sit at the premium end. Iterate at the cheap tier first — get the prompt right, get the framing right, get the source image right — then make the final pass on a flagship model. This single workflow change saves significant credits over the course of a project.

Resolution and duration are the other levers. A 1080p run can cost twice a 720p run on the same model, and an 8-second run can cost twice a 4-second run. The credit estimate on the node updates as you change resolution and duration, so you can see the trade-off live. For social and review-pass work, default to 720p and short durations; bump to 1080p and longer durations only for the final cut.

If you are stuck choosing, follow this default: Hailuo for first-pass iteration, Seedance 2 for image-to-video finals, Sora 2 or Veo 3 for cinematic finals, Kling for multi-shot narrative, Vidu or Kling for character consistency. This covers 90% of jobs. The remaining 10% — niche aesthetic preferences, very long durations, specific provider strengths — is where you experiment.