By attribute · 9 models · 105 documented failure modes
Which AI video model follows long, multi-instruction prompts?
Every model documents an instruction-drop / prompt-adherence failure that worsens as prompt length grows. Front-load must-haves and keep prompts short.
Short answer
No AI video model reliably follows long, multi-instruction prompts. Every covered model documents an instruction-drop failure that worsens as prompt length grows — camera moves and object counts get dropped first. Front-load your must-have instructions and keep prompts short, rather than expecting a single model to obey a dense prompt.
Prompt adherence degrades with length across every model because the model attends to a limited budget of instructions per generation, and the extras fall off the end. The first casualties are usually camera directions and precise counts. The fix is prompt design, not model choice: lead with the one or two things that must be right, cut adjectives that do not change the shot, and split a complex shot into separate generations you assemble later.
See the documented evidence: the failure in the catalogue, the full failure catalogue, or the overall consistency ranking.
Full context
Documented failure profile, every model
| Model | Documented modes | Holds best on | Documented weak spot |
|---|---|---|---|
| VeoGoogle Veo 3 | 13 | native audio, single-shot photoreal, lighting | long-prompt instruction drop, camera-motion-ignored on locked-off shots |
| RunwayRunway Gen-4 | 13 | character identity across cuts (Scenes mode) | hand anatomy on close-ups, prompt-ignored on dense prompts |
| SoraOpenAI Sora 2 | 12 | stylized motion (historically) | camera-control failures, multi-character interaction |
| SeedanceByteDance Seedance | 12 | short stylized clips | style-preset drift, motion drift over long clips |
| LumaLuma Dream Machine Ray-2 | 12 | lighting realism, atmospheric single takes | identity drift past ~3 cuts, camera-path drift |
| ViduVidu | 11 | reference-to-video character carry | motion plausibility, color drift |
| PikaPika 2.0 | 11 | stylized short-form, the closest Sora-style substitute | face distortion on long clips, motion failures |
| KlingKling 1.6 | 11 | human motion on simple single-subject shots | motion-blur overload, prompt adherence on complex scenes |
| HailuoHailuo MiniMax | 10 | expressive faces on close-ups | camera-shake artifacts, physics collapse |
Which model holds…
Pick by the thing that has to stay consistent
Holds a consistent face across cuts
Runway Gen-4
Runway Gen-4 Scenes mode is the only documented model built specifically to hold a character across multiple cuts; others drift after a few cuts.
Holds readable on-screen text
No model reliably does
Text rendering is a documented failure mode for every covered model — all garble past roughly six characters. Add text in post instead of relying on the model.
Holds correct hands in close-up
No model reliably does
Hand-anatomy failure is documented across every model. Frame hands away from camera or expect to re-roll; no model has solved close-up finger topology.
Holds cinematic lighting in a single take
Luma Ray-2
Luma Ray-2 documents the fewest lighting-related failures and leads on photoreal cinematic light for mood-led single-shot work.
Holds native audio with the video
Veo (Google Veo 3)
Veo is the only covered model with native audio generation; the rest produce silent video that needs separate audio.