AI video vendors and researchers sell on one ratio: seconds of video per second of compute. "Faster than real time" is a powerful line in a pitch deck. It is also, almost always, measured on a data-centre GPU that costs more than the workstation your team will actually use.
If you are evaluating in-house AI video, the claim you need is different: which model fits which job, on the hardware you have.
Three Open Models, Three Different Jobs
LTX Video generates video from text. Its research paper reports 768 by 512 video at 24 fps and five seconds of video in about two seconds, measured on an NVIDIA H100.
Stable Video Diffusion animates a still image. Its paper credits data curation and a three-stage training process for its quality.
AnimateDiff adds a motion module, trained once and plugged into personalised image models, so a brand-styled image model can animate without retraining.
Match the Model to the Asset
| Business need | Model | Memory needed (catalog) |
|---|---|---|
| Animate an approved still, such as a product shot | Stable Video Diffusion | 8.0 GB |
| Explore concepts quickly from a brief | LTX Video | 9.4 GB |
| Keep a house style consistent in motion | AnimateDiff | 8.0 GB |
Valve's September 2026 Steam Hardware Survey shows 16 GB as the most common consumer VRAM tier at 27.21%. All three models fit on such a card on paper. On 8 GB cards, still the second most common tier, they sit at or over the line and will lean on memory offload, which costs speed.
Plan Around the Real Constraints
- Expect render time to exceed clip time on workstations. The H100 figure is a ceiling, not a service level.
- Treat clip length as a budget. Frames equal frame rate times duration, and each frame costs compute.
- Generate short, edit long. Short clips are cheaper to retry and easier to approve. Sequences are built in the edit.
- Keep the assembly in-house. The time sink is rarely the generation; it is cutting clips into something publishable.
Where the Value Is Created
For most teams, a usable AI video capability is less about the model than the pipeline around it. Vision Studio, our open-source app, pairs the three models above with a storyboard that ghosts neighbouring scenes for continuity and a timeline that exports finished MP4 with transitions and AAC audio. LoRA styles apply to video as well as images, which keeps brand look consistent across both.
Why the Research Figure and the Desk Figure Differ
A paper's speed figure is a precise answer to a narrow question: how fast does this model run on this accelerator, at this resolution, in a controlled setting? That is useful for comparing models with each other. It says little about a creative team's day, which runs on mixed hardware, mixed clip lengths, and a lot of iteration.
The difference is not a flaw in the research. Data-centre accelerators have more memory, more memory bandwidth, and more compute than workstation cards. When a model fits comfortably, it runs at full speed. When it only fits by moving parts of itself in and out of memory, every step pays for the trip. For workstation planning, the questions are whether the model fits and how much headroom remains, not the headline ratio.
Build Versus Buy Is a Volume Question
Hosted AI video services remove the hardware question by charging per use. In-house generation removes the per-use charge by requiring the hardware and the skills. Neither is right in general. The deciding factors are predictable:
| Factor | Favours hosted | Favours in-house |
|---|---|---|
| Volume | Occasional clips | Steady, high output |
| Confidentiality | Public-facing, unrestricted material | Unreleased products, client material |
| Style control | Generic looks are acceptable | A house style must hold across assets |
| Hardware already owned | Machines below the memory line | 16 GB or more already in the fleet |
Most organisations end up with both: hosted tools for one-off experiments, and an in-house capability for the steady, confidential or brand-critical work.
Govern the Output, Not Just the Tool
Video raises the same governance questions as images, at higher stakes. Each model carries its own licence. Approved clips should be stored with the inputs that shaped them: the source still, the prompt, the style references. And generated footage that depicts real people or products needs the same review path as any other marketing asset. Settle those rules before the first clip ships, not after.
Questions for Any AI Video Vendor
- What hardware produced the speed figures, and what does the same job take on a workstation with 12 or 16 GB of memory?
- What is the maximum clip length and resolution before quality or memory becomes a problem?
- Which models run underneath, and what licence does each carry?
- Where do prompts, source images and outputs go, and are they used to train anything?
- How are clips assembled and exported, and can the output drop into the editing tools the team already uses?
A vendor who answers these specifically is selling a capability. One who answers with the headline ratio is selling a demo.
Start With One Use Case
The fastest way to learn what AI video can do for your organisation is a contained pilot. Pick one recurring need, such as animating product stills for social posts, and run it for a month on hardware you already own. Measure time per approved clip, the share of generations used, and the editing time around them. Those three numbers will tell you more about whether to invest further than any vendor demonstration, and they will be your numbers, on your work.
The Evaluation Checklist
- Ask for results on hardware you own, at the clip lengths you need.
- Map each content type to the model built for it, rather than one model for everything.
- Budget for memory, not just compute.
- Evaluate the editing and export path, not only the generator.
The full hands-on guide, with clip budgets and workflow is on the Vision Studio blog at vision-studio-x.com/blog/local-ai-video-honest-numbers. Vision Studio is a free, open-source, local-first AI generation app built by Strategia-X. If you want a second pair of eyes on how your organisation adopts AI tooling, talk to us at strategia-x.com.
-Rocky
#AIVideo #TechnologyEvaluation #ContentOperations #GenerativeAI #LocalAI #OpenSource #StrategiaX #VisionStudio