We set out to test one AI video vendor. We came back with a fully automated, compliance-gated, brand-owned production system — three production lanes, an owned spokesperson, and every learning documented for the agent team. This is the whole journey, every video included.
Same compliance-locked script throughout — the only thing that changes is the machine behind it. Every video below is real output, in chronological order. Press play on any of them.
Day one: what can we make with zero budget, then with the incumbent avatar vendor?
Kling 3.0 and Seedance 2.0 via API — photoreal B-roll under a premium voice.
An AI spokesperson we own — one portrait, animated and lip-synced two different ways.
Real brand assets, seam-hiding edits, and seven controlled virality experiments.
One body, many skins — and a faceless lane nobody else in the niche is running.
“TCPWOMAN” is not a vendor avatar. She's a trained identity model we control — $2.95 to create, infinitely reusable, consistent across four independent mechanisms. No licensing, no usage rights, no vendor lock-in.
ONE SOURCE PORTRAIT → LORA MODEL (16 TRAINING STILLS) → B-ROLL SCENES → SINGLE-GENERATION MULTI-SCENE SEQUENCES. SAME FACE, EVERY FRAME, VERIFIED.
Not vendor marketing — measured findings, each one paid for with a controlled experiment.
Seedance produced the most photoreal presenter we saw — and we rejected it. It invents a new person every generation, its policy wall means she can never be made consistent, and its auto-voice garbled the one legal term that matters. An owned, growable identity wins on brand equity, compliance, and compounding value.
Seedance blocks all photoreal faces in reference images. Kling 3.0 has no such wall — its elements feature accepted our spokeswoman and held her identity across a three-scene single generation. Knowing which engine permits what is as valuable as knowing which is prettiest.
Seven experiments against a neuro-prediction scorer: sustain scored a perfect 100 every time, while no first-3-seconds treatment moved the hook score more than ±3. The lesson: our content holds viewers; opening-frame gimmicks don't move the needle — front-loading the strongest content might. And: never remove the human from the cold open (scores dropped).
“The lipsync feels off” became a silence-map diff proving the restyle vendor inserts 0.2–0.6s gaps mid-sentence at caption boundaries. We built an automatic repair (tighten_gaps.py) that cuts inside the silences — lipsync-safe, and it recovers 4–8% runtime. Every anomaly this session got the same treatment: measured, root-caused, then fixed in the pipeline permanently.
Competitor analysis of 9 top credit Shorts found they win with claims we legally cannot make (guaranteed deletions, “legal loopholes”). Our answer: a hard compliance gate in the pipeline (verbatim script checks, mandatory no-guarantee line, disclaimer end card) — and a content strategy that turns their violations into our debunk material.
Higgsfield resells the same engines we hit directly — its real value is three exclusives (Wan avatars, the virality scorer, style restyling) and its failure modes are real (broken MCP video path, opaque errors, plan-gated features). We use it surgically and keep the backbone on direct, pay-per-use APIs with no automation policing.
One skill invocation runs this end-to-end. Every step is scripted, every credential vaulted, every gotcha documented in the team knowledge base.
| Role | Engine | Why it won |
|---|---|---|
| Talking avatar | Wan 2.7 | Only engine that animates our identity from our audio; most expressive delivery |
| Identity persistence | Owned LoRA + Kling elements | Same face across stills, B-roll, and multi-scene generations — model-enforced |
| Cinematic B-roll | Seedance 2.0 | Best image quality tested, native 1080p |
| Hooks & transitions | Kling 3.0 Pro | Identity-consistent push-ins; start+end frame interpolation = guaranteed beats |
| Voice | MiniMax speech-02-hd | Natural professional read at ~$0.10 per video |
| Restyling | Shorts Studio + custom brand preset | One body → clay / motion-graphics skins; preset learned our brand for free |
| Scoring | Virality Predictor | Free in beta; used as regression check, not KPI (we measured its limits) |
| Assembly & captions | Local pipeline (ours) | $0, word-exact verbatim captions, brand system, disclaimer end card — full control |
Traditional UGC creators charge $150–500 per video. Ours, at final-cut quality, per finished ~40-second video:
TCPWOMAN presenting to camera with cinematic cutaways of herself. Trust-building educational content. The flagship lane.
Faceless editorial collage with the stamp device. Zero identity risk, maximally distinctive, cheapest to scale. Built for the debunk series.
Any finished video re-skinned (clay, brand motion-graphics) for platform variety — same message, new scroll-stopper, no reshoot.
“Every video in this deck, every experiment behind it, and the spokesperson herself — less than the cost of one traditional UGC video.”
The machine is built. What we point it at is a marketing decision.