Overview & how to read this

This is the playbook for building your private video automation studio — realistic meme and promo videos with consistent characters, captions, voice, and sound, without burning money. Every capability below lists the best route for quality at low cost, the actual models/tools, what it costs, and — honestly — where it breaks.

Two rules shape everything:

• One capability at a time. Get it working, then add the next. Don't build everything at once.
• Play to strengths. Simple, restrained motion and strong starting images work well today. Complex object physics and exact action-replication are the hard, unreliable, expensive parts. We lean into the first and treat the second as careful experiments.

The roadmap

Attack in this order. Each phase builds the foundation for the next, and the harder stuff comes only after the cheap wins are solid.

Cost cheat sheet

Observed prices from our actual jobs and current listings. Listed rates are not always the invoice for your exact settings — treat these as planning numbers. No premium models or auto-retries are used.

1. Consistent characters / avatars

Medium   The same person appearing across every video.

Best low-cost route

• Use a fixed, approved character asset — a set of reference images (face, hair, body, clothing) that you generate once, visually approve, and reuse everywhere. This is what made our four-scene test use one woman across shopping, baking, and grill clips.
• Reference-image conditioning (pass the approved face/character image into each generation) is the cheap starting point. Nano Banana 2 for stills, then feed the same reference into the video.
• Multi-reference (WanAnimatePlus, 3 references recommended) only if identity drifts between shots.

What to avoid (for now)

Practical build

• One approved character photo → reuse as the identity reference in every job.
• Keep face, hair, build, clothing stable; make profile/back views only if useful (and approve them — generated extra views can contradict the identity).
• Add a "characters" section to the dashboard: save each approved character, click to reuse.

Cost: ~$0.027 per character still. Reference reuse is free.

2. Talking head / "yap" videos (8–60s)

Medium   A person talking to camera, short form.

Two routes — pick by what you already have

Attack order (from research)

1. Approve the audio first. Get the voice right (pronunciation, pacing, breaths, natural pauses) before any video. This is cheap and gates everything.
2. Test LongCat Avatar 1.5 at a low-cost setting (photo + approved audio). Compare InfiniteTalk if needed.
3. If your movement footage is already good, use MuseTalk / LatentSync as a mouth-only finishing pass (cheaper — reuses the body).
4. HeyGen Avatar IV only as a quality fallback at the approved budget ceiling.

Cost: voice ~$0.03–0.04/1,000 chars + video. 8s test ≈ $0.10–0.30.

3. Timelapse / contractor transformations

High risk (if pure AI)   Builds, cleanups, demolitions, interiors, movers. Transformational, stable environment (garage, yard).

Best low-cost route — don't let AI "do" the task

• Use real footage of the real action (or stock), then edit deterministically. AI only for bounded appearance changes (face, clothes) — not for the physical removal itself.
• Fixed camera (locked tripod) so the environment stays stable. Keep an object "ledger" of what starts where and exits when.
• Hard cuts between stages (before → mid → after) with honest time-sampling. A real cut is honest; a dissolve of an opaque object is not.
• Worker exits, then hold on an approved clean still for the last second (free — no model call).

When you want a pure-AI transition (optional)

Wan 2.2 start-to-end (first/last frame) can blend a messy → clean still. Good for a stylized "before/after" reveal, not for claiming continuous physical cleanup. Treat it as a visual effect, label it as such.

Garage test: 2 stills ($0.027 each) + 8s start/end video ($0.16) = $0.214, but not production-ready. Real-footage editing is cheaper and honest.

4. One person doing real movements

Medium–High   Pushing a cart, cooking at a grill, cooking at a stove, walking down the street filming herself.

Best route — real action, swap the person

• Shoot (or use stock of) the actual action, then human-replace the actor with your character using Wan-Animate "Mix" mode. This keeps the real cart, tray, grill, hands, and contacts while changing who's doing it.
• Preserve the props and contacts — the grill lid, tray, cart handle must stay real and correctly gripped. AI regeneration of these is what breaks.
• For a simple version without a swap: just use real footage + captions.

What we saw

Per-action notes

Wan 2.2 video ~$0.16/8s. Human-replacement via custom workflow: request a quote before running.

5. Podcast — one or two people, natural interaction

Medium   People talking and reacting to each other.

Best low-cost route — audio first, then edit shots over it

1. Master one continuous audio timeline first (both voices, natural pacing). Keep each speaker as a separate stem.
2. Build the video as edited closeups over that audio — shot/reverse-shot. When person A talks, show A; reuse brief silent "listening" shots of B (blinking, nodding).
3. Only after that works, try a true two-person wide shot with both talking (harder).

Tools

• Voices: Qwen3-TTS (two-person dialogue), or separate TTS per speaker (see Voices below).
• Video: InfiniteTalk / LongCat (single person), MultiTalk (true two-person, riskier). MuseTalk/LatentSync for lip-sync finishing.

Voice cheap; video 8–15s turns first. Build 60s from accepted turns.

6. Mimic / replicate a video + human replacement

High risk (complex props)   Take a video (e.g. a woman jump-roping) and replace her with any character — real or imaginary.

The route

• Wan-Animate "Mix" (replacement) mode is exactly designed for this: it keeps the real scene, motion, and objects while swapping the person's appearance. This is the one to use for "replace a person in an existing video."
• Animation vs. replacement: animation makes a new shot; replacement edits the person inside the existing shot. For mimic, you want replacement.
• Imaginary characters: humanoid characters (fantasy people, mascots) map onto human motion well. Radically different body plans are high-risk.

The jump-rope caveat

Our earlier yoga mimic test (Kling motion-control) kept the new character but created ghosted duplicates around the crossfade edits and changed a pose. Lesson: use continuous single-shot references, not edited montages with dissolves.

Request a quote for the Wan-Animate workflow before running; no premium fallback.

Voices & music

Cross-cutting — every talking, podcast, or promo video needs these. Voices were called out as important, so get them right early.

Text-to-speech (cheap)

Background music & sound effects

• Keep a small library (ambient, whoosh, pop, click). Mix in locally with FFmpeg — free, and you can swap the music without regenerating video.
• Add "ducking" (music dips under speech).
• Some video models generate sound natively, but it's less controllable and costs more — keep sound as a local add-on.

The interface plan

Everything lives in one private dashboard (already started at vstudio.viewervelocity.com). It's how you run each capability and see what you've made.

• Jobs dashboard — every job saved as a card: starting image + prompt + output + cost. Filter and history.
• "Re-run with new prompt" button per job — the cheap-revision workflow (fix a deformed arm without starting over).
• Finishing step — captions, motion graphics, background music, and voice in one place.
• Characters & voices libraries — save approved characters and voices, reuse across videos.
• Presets — saved prompt templates per content type (product promo, meme, talking head, etc.).

Glossary