AI video, stock footage or filming it yourself: how to choose, shot by shot

Most guides on this topic pick a winner. I think that is the wrong question. The right question is: what must this shot do? A shot that proves something needs a camera. A shot that sets a mood can come from a stock library or an AI model. This page gives you a rule, a table, a shot list template and a test you can run in a week.

I will use Higgsfield as the AI example. What I say about it comes from its own pages and terms, and I tell you when something is a vendor claim.

This page may contain affiliate links from which Blur Studio earns commissions.

Quick answer: choose by what the shot has to prove

Ask two questions about every shot in your plan:

1. Would a viewer feel cheated if they learned this shot was generated? If yes, film it or use real stock. If no, AI is fine.
2. Is this shot making a claim? A claim is anything like "it works", "it fits", "it looks like this", "it happened here". A claim needs a camera.

That is most of the decision. The rest of this page is detail.

The three options in plain words

• AI video. You write a prompt (a written instruction) or give a photo, and a model makes a short clip. Higgsfield hosts many such models from different companies, and its price table lists most clips at a few seconds each.
• Stock footage. Clips shot by other people and licensed to many buyers. It can be paid (a library subscription or per clip) or free. Pexels, a free stock photo and video site, is the free lane.
• Filming yourself. A phone, a camera or a local videographer you hire. You control exactly what is on screen.

Decision table

Question	AI video	Stock footage	Filming
Cost per clip	Low per try, but several tries per accepted clip plus your time	Free to a licence fee per clip or a library plan	Highest for one clip, lowest per clip if you shoot many in one day
Speed	Minutes per try	Minutes to search and download	Planning, a shoot day and edit time
Looks unique to you	Specific to your prompt, but outputs may not be unique across users	Others can license the same clip	Yours alone
Realism of real places and people	Weak on faces, hands, text, crowds	Strong, it is real	Strongest
Control over the exact scene	Medium, with many retries	Low, you take what exists	High
Product accuracy	Risky	Not your product	Exact
Legal and rights checks	Tool terms, likeness, brands	Licence terms	Releases for people and places
Best for scale and variants	Yes	Some	No

The evidence shot vs atmosphere shot rule

Tag every shot in your storyboard with one letter:

• E for evidence. Product in use, your team, your premises, results, a customer speaking. These go to a camera.
• A for atmosphere. Mood, transitions, abstract backgrounds, scene setting, a generic street at dawn. These can come from AI or stock.

Only A shots go to AI or stock. This one habit prevents most trouble, from product mix ups to complaints that an ad was misleading.

Five common jobs and the lane for each

1. A product ad for a physical item. Film the product (evidence). Use AI for a stylised background or an opening mood shot (atmosphere). If you want to animate your real product photo, approve a still first, then animate it, and check the result against the real item.
2. A service business explainer. Film the owner and the work. Use stock for generic office or street b-roll (supporting footage that covers a voiceover; see adding cutaways to a talking-head video). Use AI only for an abstract intro.
3. A social media hook. AI is a good fit: eye-catching, stylised, fast to test in several versions. Nothing in the hook should claim anything about your product.
4. A launch teaser with a made-up scene. Something like "a city of the future". Stock cannot supply it, filming is too costly, so AI wins.
5. A testimonial. Filming only. A generated "customer" is a fake review, and I would not do it. Consent and likeness issues apply too; see the photo and media release form.

Costs with numbers you can replace

These figures are made up to show the method. They are not prices from any vendor. Replace each assumption.

A 30 second ad needs 8 shots. For a full first-ad plan, see your first AI video ad.

• Stock: 8 shots x an assumed $30 per licence = $240. Add search time, say 1 hour. For a real library price, Storyblocks lists its Essentials plan at US$21 a month billed annually, with unlimited stock downloads under an individual licence (read 2 November 2026; check its page for current plans and for the business licence you may need for client work).
• AI: 8 accepted shots x an assumed 4 tries each = 32 tries. Multiply by the credit cost per try for the model you pick (the Higgsfield price table shows it per model and resolution; check the live page). Add 90 minutes of your time.
• Filmed: a half day with a local videographer. Get a real quote; do not rely on a number from a blog.

A vendor, Genra (an AI video company), lists stock clips at $15 to $500 each and AI clips at $0.50 to $5 each on its comparison page. It is one vendor's estimate with no named dataset, so I am only mentioning it as a vendor claim. Collect your own numbers.

A free lane exists. Pexels says all its photos and videos are free to use without attribution, and you may change them. It does not allow showing identifiable people in a bad light, implying endorsement, selling unaltered copies or using a clip as part of a trademark or business name. That is useful for generic b-roll. It does not give you an exclusive look, because anyone can use the same clip.

Rights and licence checks for each lane

• Stock: read the licence. Free libraries often forbid implying endorsement. Paid ones may limit editorial use or large ad impressions. Keep the receipt or the licence page for each clip.
• AI: read the tool's terms. Higgsfield's terms say it does not claim ownership of your inputs or outputs and does not restrict commercial use of outputs. They also say outputs may not be unique, that Higgsfield may use content to train and improve its models (Enterprise customers excluded), and that you are responsible for clearing rights to likeness, brands and anything you upload. Some prompts get blocked, such as copyrighted characters and brand names. The free tier adds a visible watermark, and Higgsfield's pricing table marks commercial use as not included on the Free plan even though its Terms say they do not restrict it, so use a paid plan for client work or ask Higgsfield support first. Never say "you own it and it is unique". Read is Higgsfield safe and legit and, for voice tools, ElevenLabs commercial use as an example of reading terms before use.
• Filming: get releases from people who appear and permission for private places. Music is licensed separately.
• Disclosure: platforms and some laws ask for labels on AI content, and Higgsfield's terms say you must disclose AI where law requires and must not remove provenance signals (labels that show a file was made by AI). Check each platform's current rules and your local law. This is information, not legal advice.
• Real people: any recognisable person, including a lookalike prompt or a generated version of a staff member, needs a release. This guide stays with generic scenes. For real faces and voices, see the lip sync and talking photo guide.

How to buy stock without it looking like stock

Most pages on this topic treat stock as one flat option. It is not. These habits make stock clips look chosen, not grabbed:

• Search for actions, not topics. "Hands pouring coffee into a mug" finds more usable shots than "coffee business".
• Run the generic test. If the clip could sit in a competitor's video without anyone noticing, skip it. Pick clips with something specific in them, such as a local landmark, a real tool of your trade or an unusual angle.
• Check the same clip is not everywhere. Search a few words from the clip title on a video site. If ten brands use it, so can your viewer's memory of them.
• Match the licence to the use. A clip licensed for a website may not cover a paid ad, a TV spot or a client resale. Read the line for your exact use before you download, and save the licence page.
• Look for releases on people. A clip with a clear face should say the person signed a release (written permission). If it does not say, choose another.
• Grade to match. Stock from different makers has different colour and sharpness. Use the same colour treatment across all clips so the cut feels like one film.

Filming yourself: a one hour phone checklist

Filming feels like the expensive lane, but a short shoot with a phone is cheap if you plan it. Before you start, write down five shots that only you can show: the front of your premises, you or your team at work, the product in real use, a close-up of the key detail, and one wide establishing shot. Then:

1. Film horizontally for a website and vertically for social, or frame wide enough to crop both ways.
2. Shoot near a window, not under ceiling lights. Wipe the lens first.
3. Hold each shot for 10 seconds, even if you will use 3. Editing needs spare footage.
4. Record 30 seconds of room tone (the quiet sound of the place) so an editor can smooth cuts.
5. Get signed releases from anyone who is recognisable, on the day.

One hour of this can supply real evidence shots for months. Use AI or stock only to fill the gaps around them.

Mixing the three: a sample shot list

Copy this into a sheet. One row per shot.

#	Shot	E or A	Source	Max tries	Check
1	Opening mood: morning light on a street	A	AI or stock	5	No brand or text visible
2	Owner says the promise to camera	E	Filmed	n/a	Release signed
3	Product in use on the bench	E	Filmed	n/a	Matches the real product
4	Transition: abstract shapes	A	AI	5	Smooth loop
5	Generic busy office	A	Stock	n/a	Licence saved

A cheap trick: book a "raw footage day". A local videographer, or you with a phone, films every E shot for the whole quarter in one half day. After that you fill A shots with AI or stock. That is usually cheaper than commissioning each video separately.

Save a "provenance sheet" in the project folder: for each AI shot, keep the prompt, the tool, the date and the original frame. If a client or platform asks later, you can answer in minutes.

A waste-saver for AI shots: approve a still image first, then animate it, and cap tries at 5 per shot. When the cap is reached, switch to stock or filming. Count "accepted shots per 10 tries" in your test; that is the number your cost math needs.

To read about credit cost and how to avoid wasting tries, see how to spend Higgsfield credits wisely.

A 7 day test with pass and fail marks

1. Day 1. Pick one real brief and write the shot list with E and A tags.
2. Day 2. Find 8 stock candidates for the A shots. The free lane counts.
3. Day 3 and 4. Generate the same A shots with AI, with a cap of 5 tries each.
4. Day 5. Film the E shots, even on a phone.
5. Day 6. Assemble three versions: all stock, hybrid, and AI heavy.
6. Day 7. Blind review by two colleagues and one customer.

Pass for the AI lane: at least 6 of 8 shots accepted within the cap, and no reviewer flags a product inaccuracy or says "this feels fake". Fail otherwise: use stock or film those shots.

When AI video is the wrong choice

• The shot shows your real product, team, premises or results.
• You need generic real-world establishing shots (a real city, a real crowd). Stock is cheaper and looks real.
• You need a person to speak a claim, such as a testimonial.
• You work in a regulated field where a visual is itself a promise.
• You need a long continuous take. AI clips are short.

Stock beats AI for generic real-world b-roll. Filming beats both for evidence. Specialist video models from other companies, such as Kling (from Kuaishou) or Veo (from Google), may beat any single choice for certain motion; Higgsfield's pitch is that it offers several in one subscription, which helps if you want to test them side by side. See Higgsfield vs Runway for one comparison.

How I'd approach this

I would run the shot list first and let the tags make the decision. Film the evidence once. Fill the atmosphere with free or paid stock when it is generic, and with AI when I need something that does not exist. I would try one month on a paid AI plan only after the 7 day test passes, and I would check the cancel and refund rules before paying (Higgsfield says refunds are limited to 7 days with no credits used, and renewals are not refundable). You can see the current plans on the Higgsfield pricing page, since prices and promotions change.

If you make videos often, Blur Studio can build the repeatable workflow around this: a shot list template, a shared asset library and an approval page for clients. See how I work with businesses.