Bonus stage: AI video

AI video should be free too.

Gurus are now charging $497 to teach you to type a sentence into a video generator. Here is the whole thing: four tools, what they really cost, how to use each one, and the rules that separate scroll-stopping video from AI slop.

Prices checked 2 October 2026. AI tools change pricing constantly, so confirm on each site before you pay.

V.1 Pick your tool V.2 The real cost V.3 How to use each V.4 The free game V.5 Seedance & Wan prompts V.6 Your stack by budget

V.1

Four tools. Four different jobs.

These are not four versions of the same thing. Pick wrong and you burn credits fighting the tool. Pick right and a 30-second ad takes an afternoon.

Start here

Dreamina

CapCut's AI image and video generator. Free daily credits, so you can learn without paying.

Best for: Beginners, social clips, testing ideas
Cost: Free daily credits. Paid plans vary by region.
Skill: Beginner

Most models

OpenArt

One account, many image and video models. Strong for consistent characters across scenes.

Best for: Storytelling, recurring characters, model shopping
Cost: From $14/mo. 40 free trial credits.
Skill: Beginner to intermediate

Cinematic

Higgsfield

Built for camera moves and cinematic shots. Gives you access to many top video models in one place.

Best for: Ads, product hero shots, dramatic camera work
Cost: From $15/mo. No real free plan.
Skill: Intermediate

Code-made

Claude Code

Does not generate footage. It writes code that renders video: motion graphics, animated text, data, explainers.

Best for: Branded explainers, captions, templates you reuse forever
Cost: Included with Claude Pro, $20/mo
Skill: Intermediate (no coding needed, but you will run commands)

V.2

The real cost. Not the headline price.

Subscription price is the wrong number to look at. The real number is cost per usable clip, and that depends on how many attempts it takes to get one you would actually post.

Tool	Free option	Entry plan	Higher plans	Watch out for
Dreamina	Daily free credits (around 120 a day as of Aug 2026)	Basic, Standard and Advanced tiers; price shown in your account and varies by region	Same as left	Cost per clip changes with length, resolution, audio and model. Check the credit cost before you hit generate.
OpenArt	40 credits, one time, expire after 7 days	Starter $14/mo: 4,000 credits (OpenArt estimates about 50 videos)	Plus $34/mo (12,000), Pro $56/mo (24,000), Wonder $240/mo (106,000). Discounts often shown on annual billing.	Unused credits do not roll over. Premium video models eat credits much faster than the "50 videos" estimate.
Higgsfield	Trial credits only	Starter $15/mo: about 200 credits	Plus $39/mo annual or $49 monthly (about 1,000 credits). Ultra $99/mo annual or $129 monthly (3,000+).	Credits do not roll over. Independent reviews put real cost around $2 to $4 per short clip on Plus.
Claude Code	Not on the free Claude plan	Claude Pro $20/mo ($17/mo billed yearly)	Max from $100/mo for heavy use	Shares usage limits with your normal Claude chats. Rendering happens on your own computer, so no per-clip fee.

Sources: claude.com/pricing, openart.ai/pricing, dreamina.capcut.com, and independent Higgsfield and OpenArt pricing breakdowns, all checked late September to 2 October 2026.

What a 30-second ad really costs

A 30-second ad is about six 5-second shots. Plan on 3 attempts per shot to get one keeper. That is 18 generations.

Dreamina
Spread the 18 generations across a few days of free credits. Slow, but free.

~$5+

OpenArt
Using cheaper models on Starter. Premium models can cost several times more.

$36 to $72

Higgsfield
At about $2 to $4 per clip in credits. Best-looking cinematic shots, highest cost per keeper.

$20/mo

Claude Code
Flat. Unlimited renders of motion graphics. No real-world footage.

The biggest cost is not the plan. It is wasted generations. Every rule in V.4 exists to cut your attempts per keeper from 5 down to 2.

V.3

How to use each one.

The workflow that works in every tool: make the image first, then animate it. Images are cheap and fast to iterate. Video is not.

Dreamina

Free clips in 15 minutes

Sign in at dreamina.capcut.com with your CapCut, TikTok or Google account.
Open the image generator. Make your opening frame first using the prompt formula in V.4.
Pick the best image. Choose "image to video". Describe only the motion and the camera, not the scene again.
Generate at the shortest length first. Only extend or upscale a clip you are keeping.
Download, then cut, caption and add music in CapCut.

OpenArt

Same character, every scene

Use the 40 trial credits to test before you subscribe.
Create a consistent character from 1 to 3 reference images of your person, mascot or product.
Generate each scene as a still with that character. Fix faces and outfits now, not after animating.
Animate each still with an image-to-video model. Compare two models on one shot before committing a whole project.
Use the cheapest model that looks good enough. Save premium models for the hero shot.

Higgsfield

Shots that look expensive

Start with a strong still: your product photo or a generated keyframe.
Pick a camera move preset (dolly in, orbit, crash zoom). One move per shot. Never stack them.
Write a short motion prompt: what moves, how fast, what stays still.
Run one test at the lowest cost setting. Only rerun in high quality once the motion is right.
Use it for the 2 or 3 hero shots of an ad. Fill the rest with cheaper tools.

Claude Code

Videos made from code

Get Claude Pro, install Claude Code, and open it in an empty folder.
Ask it to set up a Remotion project (a free React video framework for individuals and small teams) and install ffmpeg.
Describe the video in plain English. Copy the prompt below.
Preview in the browser, ask for changes in plain English, then render to MP4.
Save it as a template. Next week, swap the text and render a new video in minutes.

Build a 20-second vertical (1080x1920) Remotion video for [brand].
Scenes:
1. (0-3s) Hook text: "[bold claim]" slams in, shakes once.
2. (3-12s) Three benefit lines appear one by one with icons.
3. (12-17s) A number counts up from 0 to [stat].
4. (17-20s) Logo + CTA: "[call to action]".
Style: black background, hot pink #FF007F and yellow #FFD700, Playfair Display headlines.
Add burned-in captions. Keep every text block readable for 1.5s minimum.
Then render it to out/video.mp4.

V.4

The free game.

This is the part the courses stretch into six modules. Ten rules. Follow them and your attempts per keeper drop, your videos stop looking like AI slop, and people watch past the first second.

The prompt formula

[SHOT TYPE] of [SUBJECT] [DOING ONE ACTION], [SETTING],
[CAMERA MOVE], [LIGHTING], [STYLE / LENS], [MOOD].

Example:
Close-up of a barista pouring latte art in slow motion, small sunlit cafe,
slow dolly in, warm morning window light, shot on 35mm, shallow depth of field,
calm and cozy.

Image first, video secondLock the look in a still. Animating a bad image just gives you a bad video that cost 10x more.

One action per shot"Man walks in, sits, drinks coffee, laughs" breaks every model. Split it into four shots.

Name the cameraDolly in, orbit, handheld, static. If you do not say it, the model guesses, and guesses are random.

Name the lightGolden hour, neon, overcast, studio softbox. Lighting is the fastest way to look expensive.

Short shots winGenerate 5 seconds, not 10. Errors pile up the longer a clip runs. Real ads cut every 1 to 3 seconds anyway.

Hook in the first secondMovement, a face, or a bold line of text in frame one. No slow logo intros. Ever.

Draft cheap, finish premiumTest motion on the cheapest model and lowest resolution. Upscale only the keepers.

Keep a prompt logSave every prompt that worked, with the seed and model. Your log becomes your style guide.

Sound is half the videoAdd music, sound effects and captions in your editor. Most people watch muted, so captions are not optional.

Hide the tellsWatch hands, text and faces. Crop, cut early, or cover with captions where the AI slips. Never post a clip with warped text.

Pick the format before the tool: 9:16 for Reels, TikTok and Shorts. 1:1 or 4:5 for feed ads. 16:9 for YouTube and your website. Regenerating for a new ratio costs a full set of credits.

V.5

Every model speaks a different language.

The formula in V.4 works everywhere. But the two models most people start with, Seedance (the engine behind Dreamina's video) and Wan (available on OpenArt), have their own rules. Some of them are opposites. Ignore them and you pay for it in wasted credits.

	Seedance (Dreamina)	Wan (OpenArt)
Shot timing	Write Shot 1, Shot 2, Shot 3 in order. Do not give exact seconds. Forcing timing makes output worse.	Write a one-line summary, then numbered shots with exact ranges like `Shot 1 [0-5s]`.
References	Define each subject once from its image, then reuse the same label every time it appears.	Label every upload and its job: `@Image1 is the woman.` Wan will not guess.
Sound	Use symbols: `（music）` `<sound effect>` `{dialogue}`	Three separate sentences: voice line, sound effect, music.
Default behaviour	Can add subtitles, logos or watermarks you did not ask for.	Adds its own voice lines, music and cuts if you say nothing.
Always add	`no subtitles or text, no logo, no watermark`	`no voice lines, no background music, one continuous take` (whichever you do not want)

Seedance on Dreamina

Write it like a director's brief

Order matters: subject, action, setting, lighting, camera, style, quality, then constraints.
Act out emotion. Not "she is sad" but "head lowered, shoulders trembling, eyes welling up".
Name body parts and speed. "Slowly raises her right hand", not "raises hand".
Keep it gentle. Sprints, big jumps and violent moves break the clip. Slow, continuous motion holds up.
One camera move per shot. Push in, or pan, or track. Never all three.
4 to 5 references max. A face close-up, a full-body shot, one scene image, one motion video. More than that confuses it.
Face drifting? Use a plain headshot plus a full-body photo. Skip character sheets with multiple angles; it reads them as different people.
Unwanted subtitles? Landscape gets fewer than portrait. Generate wide, crop to vertical after.

Use the person in @Image 1 as the main character and @Image 2 as the location.
Shot 1: [Character] walks into [place], camera follows in a medium shot. Warm evening light. She pauses and takes a slow breath.
Shot 2: Close-up. She looks up and smiles slightly, says {Line of dialogue}. <soft city noise>
Shot 3: Camera slowly pulls back to a wide shot. （gentle piano music）
Cinematic, HD, natural colors, soft lighting. Face stays stable, smooth natural motion.
No subtitles or on-screen text, no logo, no watermark.

Wan on OpenArt

Build it up one layer at a time

Start tiny: subject + place + movement. "A red fox walks across a snowy field at sunrise." Then add camera, style and sound one at a time so you can see what each change did.
Give each time range one beat. If a shot feels rushed, widen that range only.
Starting from an image? Say less. Only describe what moves and what the camera does. Repeating the scene fights the image.
Want the camera still? Say fixed shot, or Wan moves it anyway.
Editing a clip: state the one change, then "everything else stays the same".
Turn off Auto Polish once your prompt is exact. It rewrites your labels and time ranges.
Draft on Standard, finish on Prime. Change one thing per variation so you know what worked.
Reusing a character? Save them once as an OpenArt Character instead of re-uploading every time.

@Image1 is the founder. @Image2 is the office.
A short story about a founder shipping her first product.
Shot 1 [0-5s]: The founder from @Image1 sits at a desk in @Image2, typing fast. Slow push in.
Shot 2 [5-10s]: Close-up on the laptop screen as a "Launch" button is clicked.
Shot 3 [10-15s]: She leans back and laughs. Fixed shot.
She says, "It's live." Relieved, quiet, American English.
A chair creaks as she leans back, quiet office.
No background music.

The mistakes that waste the most credits, on any model: describing a still picture instead of what changes over time, leaving references unlabeled, stacking camera moves, and forgetting to switch off sound, text or logos you did not ask for.

Sources: BytePlus ModelArk Seedance prompt guide and OpenArt's Wan prompt guide. Both models update often, so if results stop matching, check the latest version of each guide.

V.6

Your stack, by budget.

$0/mo

The broke founder

Dreamina free credits for clips. CapCut free for editing and captions. Post daily, learn what works, spend nothing.

~$35/mo

The serious builder

Claude Pro ($20) for reusable branded explainers and captions. OpenArt Starter ($14) for character and product clips. Dreamina free on top.

~$70 to $180/mo

Running ads every week

Higgsfield Plus or Ultra for 2 or 3 hero shots per ad. OpenArt for supporting shots. Claude Code for the end cards and captions you reuse every time.

Do not subscribe to all four in month one. Start free, find the format that gets views, then pay only for the tool that removes your bottleneck.

Want us to make them for you?

Velquiry produces AI video content and runs the ads behind it for founders who would rather be building. Same playbook as this page, done for you.

Book a free call Back to the 5 stages