When should you switch AI video models if a shot keeps failing?

Switch models after roughly 3 prompt variations and 2 seed variations on the same shot — about 5 failed attempts — or sooner if the failure is structural (anatomy collapse, multi-character contact, physics drift). Diagnose the failure type first, then route to the model built for it: Kling for physical realism and multi-shot continuity, Seedance 2.0 for reference-locked continuity, Veo for camera motion, Runway for stylized motion.

Before you switch, diagnose what's actually breaking. Most shots fail for one of four reasons, and each one points to a different next move.

Anatomy or face collapse (hands, eyes, character drift across cuts): Don't reroll the video — trace it back to the character sheet. Ask the invideo agent to inspect the sheet for the exact panel with the error, fix it there, and store the corrected sheet in context so every subsequent shot inherits the fix. In one production, the agent identified the precise grid panel containing a continuity mistake without being told where to look. Surgical edits beat slot-machine re-rolls.

Physics or multi-character contact (ropes, props, bodies touching, carries): This is the failure mode that breaks models faster than anything else. After 2-3 retries, stop prompting and bring in a physical input — act the shot out on your phone and upload the mock as a reference video, or hand-sketch the arrangement and upload the drawing for the invideo agent to feed into the image model for a fused character sheet. One production cracked a vampire-carrying-character shot only after a hand sketch went into the pipeline; another solved a stubborn POV with a phone-shot mock. If you stay in text-to-video, route to Kling or Seedance 2.0 reference-to-video, which accept character and location references simultaneously and carry context across segments — start/end frame extension and the extend feature cannot.

Camera motion drift or one-take continuity breaks: Switch to a reference-to-video workflow. Clip the final seconds of the working segment, re-upload it to the invideo agent, and let it attach that clip plus character and location references to Seedance 2.0 reference-to-video for the next continuous segment. Veo handles deliberate camera moves cleanly when motion is the spec. The invideo agent routes to whichever model fits — you don't pick a platform per shot.

Style or tone drift (plasticky skin, wrong palette, look slipping off-reference): First, check your prompt isn't dragging in a stray attachment — over-prompting with the wrong reference image produces completely incorrect output. Then strengthen the style block with explicit negatives ("not live-action, not photorealistic") and have the invideo agent read the colours and textures of your references rather than feeding illustrated images straight in. If the look still won't lock, switch the image-generation step: Recraft for skin-imperfection portraits, Nano Banana for character sheets, GPT-Image-2 for general world plates — then push to video.

The retry threshold: 3 prompt variations + 2 seed variations on the same model is the realistic ceiling before switching. Across a 3-minute episode, the documented average was 3 generations per usable shot and 17 of the final shots stitched from 2+ generations — failure and recombination are baked into the budget. Push past 5 attempts on the same model and you're paying the sunk-cost tax. "Most shots aren't one shot. Prompt → 8 tries → Frankenstein the keepers," as the team put it — meaning combine the strongest seconds from multiple generations into one composite rather than chasing a clean single take.

The image-to-video bridge: When text-to-video keeps failing on a specific shot, generate the keyframe as an image first (Recraft, Nano Banana, or GPT-Image-2 inside invideo), approve it, then push that frame into a video model with strong image-to-video — Seedance 2.0 and Kling both accept reference inputs well. Frames-first, then video is the correct production order for consistency.

Pre-assign models to shot types before you start. Power workflows route by failure profile up front: multi-shot sequences and physical realism → Kling; character-and-location continuity across segments → Seedance 2.0 reference-to-video; intentional camera moves → Veo; stylized motion → Runway. invideo holds all of these in one place, so the invideo agent routes each shot to the right model without you switching platforms. As Hridaye, invideo's creative director, put it: "The thing that made it possible wasn't prompting. It was directing. The invideo agent didn't feel like a tool — it felt like crew."

Production reality across documented runs: $315–$750 per finished minute, 2–5 day timelines, and a 25% clip-selection rate (164 generated → 41 in the final cut on the Arcane-style episode). Overgeneration is a line item, not a mistake — switch models to fix structural failures, not to chase perfection on a shot that's already 90% there.

Watch some of these to see what works for you:

The thing that made it possible wasn't prompting. It was directing. Agent One didn't feel like a tool — it felt like crew.

— Hridaye, invideo's creative director