Model, Budget,
or Privacy
One model leads the field by a distance. But the best tool is not always the right one. Budget, ambition and privacy decide that, not a leaderboard.
The question that comes up on almost every job is which AI model is best. It matters, but it does not settle the project. A model can lead on a particular benchmark or task and still be the wrong choice for a commercial. The practical decision is what the budget, creative ambition and privacy requirements can carry, and how much material the project can reasonably send to another company’s servers.
01 - The LeaderStart with the capability the job needs
For projects where its capabilities fit the brief, Seedance 2 is a strong option for video. It accepts multiple reference images, reference clips and audio in a single pass, making identity, motion and sound clearer inputs to the process. It can also help carry consistency through a sequence and make shot-level changes without reworking every frame. Those capabilities matter more than a general claim that one model wins every job. The useful question is whether they solve the part of this project that is hardest to control.
02 · When Best Isn't RightBudget and ambition rarely match on the same day
Model choice sets the cost profile of the work: capability, iteration cost, privacy requirements and where a shot needs the highest ceiling. For the wider production budget—crew, finishing, revisions and delivery—see The Total Cost of AI Production.
For much of the work, Kling is the sensible workhorse. Strong multi-shot consistency at a fraction of the top-tier cost, which is exactly what you want when a character has to be carried across a sequence on a real budget. For images,ChatGPT's image model and Google's Nano Banana are the fastest routes to a polished, directable frame. Choosing a lighter tool for a given shot is not a compromise on quality. It is a decision about where the budget does the most good, and the finishing pipeline is what closes whatever gap remains. The skill was never picking the champion. It is knowing when you need it.
03 - The Open FieldControl and cost, on your own terms
Alongside the cloud tools there is now a genuinely strong open field, and it matters for two reasons: cost and control. Running your own models means iteration is cheap, the ninth revision costs electricity and an afternoon, not another metered bill, and it means you can condition and direct a generation far more precisely than any prompt box allows. On the image side the field is deep and getting deeper: Krea, Qwen, Flux, Z-Image, and a steady stream of others. For video, open models like LTX-2 and Wan 2.2 have become properly usable for real work. Tuned with custom LoRAs trained on your own material and finished properly, they hold their own at a fraction of the running cost.
04 · PrivacyThe guarantee is not their promise. It is your machine.
Privacy is a reason to consider a local workflow before a job begins. Sending a frame to a cloud model means sending unreleased product, talent likeness or campaign material to a third party. Vendor terms, retention settings and contractual obligations need to be checked for the specific service and project. When a brief requires material to remain on controlled hardware, a local workflow removes that transfer from the process altogether.
I have had to work exactly this way. I have delivered projects built on copyrighted material under heavy NDA, where the footage legally could not pass through a closed model at all. Regardless of how good that model was. And the closed models were, in some cases, the better tool. It did not matter. When security, copyright, and data protection outrank raw capability, the requirement is absolute, not a preference. For those jobs I built custom open-source workflows that kept every frame on controlled hardware and still delivered the result the brief demanded. That is not a theoretical stance. It is a way of working I have already had to prove.
That is not paranoia; for a great deal of commercial work it is a contractual requirement. Anything under embargo, under NDA, or simply commercially sensitive should not be uploaded to a third party on trust. Using LTX-2 and Wan for video, the open image models above, and LoRAs trained on your own assets, an entire production can run on hardware you control, with nothing ever leaving the premises, and still deliver high-quality, directable output. Privacy without a quality penalty is something only the local stack can actually promise, because it is the only one that removes the question entirely.
05 - The Real AnswerMatch the tier to the project
So there is a best model, and it still is not the answer on its own. The answer is a right choice for this project: its budget, its ambition, its volume, and how sensitive its material is. A well-funded hero spot with nothing to hide can take Seedance and enjoy the ceiling. A confidential campaign, or one that has to produce a great deal at a controlled cost, is better served by an open, local stack tuned with LoRAs. Most real jobs sit in between and use a mix, the strongest tool where it earns its cost, local where control and privacy win, all reconciled in the same finishing pipeline. Choosing the model was never the skill. Reading the project, and knowing which trade protects the client best, always was.
Frequently Asked Questions
What is the best AI video model right now?
Seedance 2 leads the field by a distance on raw quality. But the strongest model and the right model for a given project are two different decisions, and the second one is what actually gets a commercial made.
If Seedance 2 is the best, why not use it on every project?
Budget and privacy. Top-tier closed models cost more per shot and send your material to someone else's servers. When the budget is tight, or the footage is under NDA, the best model on paper is often the wrong call.
How do you handle AI production under an NDA or with sensitive material?
With open-source models run locally. Krea, Qwen, Flux, Z-Image, LTX-2, Wan 2.2, inside custom workflows, so the client's material never leaves the machine. The privacy guarantee is your own hardware, not a vendor's promise.
Are open-source models good enough for commercial work?
For many briefs, yes. With the right workflow. LoRAs, control layers, and 3D guidance, an open stack holds up to broadcast finishing. You trade a little raw quality for control, lower cost, and full data protection.
How do you choose which model to use for a specific commercial?
Match the tier to the project. The budget, the ambition, the privacy requirement, and how much shot-to-shot consistency the brief needs decide the model, not which one tops the benchmarks that week.
Whether you're an agency, a brand, or hiring me directly. Send the brief, and tell me how you'll measure success. I'll tell you what's realistic, what shouldn't be done with AI, and how to get more production out of the budget you already have.
Send me the brief → See the case study →Mainly taking on projects. Open to embedded, day-rate, or full-time roles for the right production — resume here.