Back to Playlist · VidFarm Walkthrough Tutorial

Part 16 of 33Sourcing & Clipping Raws5 min read

Sourcing & clipping raws

The three paintbrushes

VidFarm doesn't burn expensive AI credits on everything. Every visual on the timeline is painted with one of a few brushes — and for bulk creation it's usually far cheaper to reach for the first two before the last.

Copies the whole guide — prompts, next steps & all — for your AI agent

Watch this chapter

The paintbrush mental model in action: the same beat sourced as a raw clip, a HyperFrames graphic, a reused asset, and finally true AI generation — cheapest first.

Key takeaways

Every shot is painted with one of three paintbrushes: raw clips, HyperFrames motion graphics, or pure AI generation.
There's a fourth, orthogonal idea — the reusable asset library: owned or generated-once-then-reused forever.
Reach for the brushes cheapest first. AI video is the priciest brush ($1–$10+), so it's the last resort, not the default.
On-screen text is always a HyperFrames graphic — never AI generation.
Every decomposed fork carries a replication harness with two plans: (A) cheap & efficient and (B) best quality.

Founder-friendly and pragmatic

VidFarm is founder-friendly and pragmatic: we do not burn expensive AI credits on everything. Think of a finished short as a painting. Every visual on the timeline — the hook shot, the b-roll, the caption, the logo sting — is painted with one of three "paintbrushes." For bulk creation it's often combinatorially cheaper to reach for the first two before the third.

Getting this mental model right is the single biggest lever on what a video costs. A caption swap and a bit of remixed footage can cost pennies; the same beat done with generated AI video can cost dollars. Same result on screen, a 100× difference on the bill. Here are the brushes, cheapest first.

How VidFarm creates videos — three approaches (Clip Library, HTML-to-video, AI generation) layered on one timeline. — Every finished video is the three brushes layered on one timeline — raw clips, HTML/JS graphics, and AI generation.

raw_clip · the workhorseBrush 1 — Raw clips

The first brush is raw clips (raw_clip): cut and remix footage from existing long-form or short-form video. This is the workhorse for scene replace. Source it in cost order:

Your own library first — pull from your /raws and /files folders (already yours, free).
Hunt new footage — mine fresh clips out of a URL or VOD with an AI hunt (vidfarm raws scan); local compute is free.
Only then generate — reach for a brush higher up the cost ladder if no clip covers the beat.

A background video plus a foreground video (greenscreen / picture-in-picture) covers most "video meme" formats with zero generation. Learn the full sourcing flow in Sourcing & clipping raws with AI .

hyperframes · cheap & deterministicBrush 2 — HyperFrames

The second brush is HTML/JS HyperFrames (hyperframes): video-from-HTML using CSS and declarative animation, anime.js/GSAP motion, animated image and text elements, and data-viz. It's cheap, deterministic, and infinitely re-themeable — change a color or a word and re-render for free.

Tip On-screen text and graphic elements are ALWAYS HyperFrames, never AI generation. Captions, titles, lower-thirds, kinetic type, charts — these are code, not pixels a model hallucinates. It's cheaper, it's crisp, and every character is exactly what you typed.

ai_gen · the last resortBrush 3 — Pure AI generation

The third brush is pure AI generation (ai_gen): AI image, video, voice, and music. It's the most expensive brush — AI video especially ($1–$10+ per clip). Save it for a beat that genuinely can't be covered by a clip or a HyperFrame: a hero shot you have no footage for, a face that has to be invented, a scene that only exists in your head.

reusable_asset · generate once, reuse foreverThe fourth idea — reusable assets

Orthogonal to the three brushes is the reusable media asset library (reusable_asset): logos, stickers, reactions, b-roll, a-roll, your brand media kit. These are either owned or AI-generated once and then reused forever — greenscreen a subject a single time, then chroma-key that overlay into video after video. So the four method values you'll see in a replication harness are raw_clip, hyperframes, reusable_asset, and ai_gen.

The three paintbrushes — broad-stroke Clip Library, mid-fine HTML-to-video, super-fine AI-generated video. — Think of them as brushes, coarse to fine: Clip Library lays down the base, HTML paints the on-screen graphics, and AI generation fills the finest details.

Illustration

cheaper↓pricier

✂️

1. Raw clipsCheapest · the workhorseCut & remix real footage you already have or hunt from the raws feed.

🎬

2. HyperFramesCheap · re-themeableMotion graphics rendered from HTML/CSS — every on-screen word lives here.

✨

3. AI generationMost expensive · last resortGenerate images, video, voice, music. Powerful, but use it sparingly.

The three paintbrushes — reach for the cheapest that tells the shot.

All video styles powered by all three approaches — talking-head, text-on-footage, explainer, cinematic b-roll, animated, and screen capture, each a different mix. — Every style is just a different mix of the three brushes — a talking-head recap leans on raw clips, an explainer on HTML graphics, an animated short on AI generation.

cheap-first vs best-qualityTwo replication harnesses

When VidFarm decomposes a template, it produces two rebuild plans so you can pick your trade-off. Both are anchored to the same viral DNA — they just spend differently.

(A) Cheap & efficient (the default) — recaption text with set_captions/set_layer_text; build bg+fg video memes; animate HTML and images with HyperFrames; reuse library media or generate-once-reuse; greenscreen; remix raw clips out of long-form. Only if genuinely needed does it reach for an AI image, and AI video/voice/music dead last.

(B) Best quality — AI video generation by default for hero scenes; storyboard with an AI image first (cheap stills lock composition and subject); then adversarially grade the result with a coding agent (Claude Code / Codex) — render, critique against the harness, iterate. VidFarm presents both plans and recommends (A) unless you've asked for premium or the budget clearly covers it.

Note Every decomposed fork materializes this as a replication-harness.json (also surfaced as editor_context.replication_harness). It carries a summary, a recommended_strategy, per-beat scene plans with a method + fallback_method + est_credits, and — crucially — a viral_dna_guard on each beat.

The load-bearing rule

Anchor every asset and paintbrush choice to the harness — the viral_dna, its emotional_punch, the editor harness, the static-vs-pivot map. Reuse and re-skin the dressing; preserve the DNA. Never slap a sticker or logo on, or swap the footage of, a load-bearing beat (the hook, the reveal, the payoff) in a way that flattens what made the original work. Cheap doesn't mean careless — it means spending your credits only where they change the outcome.

How a director applies it

Watch the reference video and break it into beats — hook / body / reveal / payoff.
For each beat, pick the cheapest brush that preserves the DNA: can you just recaption? A bg+fg clip? A HyperFrame graphic? A reused asset? Only then AI-gen.
Execute the cheap plan beat-by-beat: raw_clip via a hunt + set_layer_media; hyperframes as a CSS/anime.js layer; reusable_asset pulled from /files or generated once; ai_gen only where the beat demands it.
Honor every viral_dna_guard along the way.

Note On the free tier, you (and your AI agent) watch the reference video and decide the paintbrushes yourselves — the harness gives you the method, not a pre-computed answer. Paid VidFarm accounts get a large library of pre-decomposed viral templates, so a fork arrives already carrying its DNA, editor harness, replication harness, and per-scene annotations. More in Free vs paid & BYOK perks .

🤖 AI Agent · devcli

Get a paintbrush-by-paintbrush rebuild plan

Point a vidfarm-equipped agent at a template and it will decompose it into beats and hand you a cheap-first plan — flagging any beat where re-skinning would break the viral DNA.

Paste to your AI agent

Using the vidfarm skill, decompose this template into beats and give me a cheap-first, paintbrush-by-paintbrush rebuild plan.
For each beat tell me whether it's:
- a raw clip (raw_clip),
- an HTML HyperFrame graphic (hyperframes),
- a reusable asset (reusable_asset), or
- true AI generation (ai_gen).
Flag any beat where reskinning would break the viral DNA, and give me a rough credit estimate for the cheap plan vs the best-quality plan.

Or drive the brushes yourself with the devcli

Terminal

# pull a fork — writes the composition + .harness/ (agent-guide.md renders the 3-paintbrush plan)
vidfarm pull <forkId>

# brush 1: reuse footage you already have
vidfarm raws search "confused reaction after reading a message"

# brush 1: hunt new footage (free local compute)
vidfarm raws scan <url|file> --prompt "..."

# brush 3: the last resort
vidfarm generate image --prompt "..."
vidfarm generate video --prompt "..."

Where to go next

Now that you know how VidFarm thinks about sourcing, learn where all that footage lives and how to grow the library:

Paint cheap, preserve the DNA

The whole game is reaching for the cheapest brush that still keeps the video working. Fork a template, read its replication harness, and rebuild it beat-by-beat — cloud renders start at $0.01.

Open the Discover feed