Unique Research · Interview · August 14, 2024 · Guest: Zhu Liang (祝亮), Co-founder & COO, Vchoo.ai · Interviewer/editor: Qiuping
After Sora, AI video is rewriting content workflows. Vchoo.ai positions as a one-stop AI story studio: one sentence → full story video in chosen style, with controllable characters/scenes. Zhu on empowering users as storytellers, not tech spectators.
Interview highlights below.
Q: Core product and scenes?
Zhu: One-stop AI story platform — text stories → shorts; web-novel promos; kids/English picture books; game trailers. Product, GTM, research in parallel since prep in late-2023 (broadcast, culture-tourism, games, comics research). v1.0 April focused story-language → shot-language (expressions/actions); 1.1 live; 1.2 (July) validated continuity of people/background/makeup; video model training targeted August; 1.3 September for stronger creation + per-shot control.
Q: AI short-drama wave?
Zhu: Tried it last year — still a gimmick; doesn’t solve creative core. Video’s job is story + emotional resonance; industry stuck generating raw shots, far from telling stories well.
Q: Uncanny / over-polished faces — when more real?
Zhu: Core tech challenge; expect multi-year quality jumps. Origin story: Chinese-culture videos for overseas — live-action, VFX, Unreal, short-drama, interactive branches all hit cost/quality/skill walls (director, writer, shoot, composite…). Evaluated tools: fragmented multimodal collab, weak controllability. Pain → build ourselves. Thesis: shots are raw material; outline, cast, scenes, boards, edit are the process — AI should democratize the full process, not only camera tricks. Continuity across shots is our training focus; don’t over-index on today’s look — give tech time.
Q: Homogeneous ideas?
Zhu: Funnel of ideas — mass similar at bottom, wild at top. Hard part isn’t CapCut vs Premiere; it’s decomposing intent into shots. Product: one sentence → inferred shots (expression, action, scene) editable without Prompt literacy; converge for coherence. Path like film→phone camera: everyone creates. “One user, one idea, one story.”
Q: Producer–consumer AI platform?
Zhu: Roadmap 1.0 infrastructure → 2.0 sharing → 3.0 produce+consume. Youth want identity content, not only algo-fed others’ posts. Self-media already blurs producer/consumer; lower tools → mesh of niche circles. Douyin/Kuaishou also investing; still exploratory.
Q: How do foundation models change pro film workflow?
Zhu: AI makes raw shots; film is process/PM — conflict. Trends: (1) new roles — AI director, AI edit trainers/interpreters; edit-before-shoot with AI animatics; (2) lower financing risk — full AI sample vs actor teasers; (3) fine-tune packs for wardrobe/props/sets reusable across sequels; open packs to fans like game mod kits. 2D→3D game leap as analogy — conditions better now; breakthroughs need 3D + temporal physics, ~1 year for many teams on raw materials. We won’t race BigCos on foundation models — app layer, continuity, style packs. Dataset quality > cost; AI pre-screen + human audit SOP; first style months, later 1–2 months. Train on skeleton + abstracted expressions (smile subtypes), not whole frames; lip-sync via mature open source; focus unsolved interaction styles.
Q: Reduce human SOP variance? H2 plans?
Zhu: TPMF (tech–product–market) loop; constant research; expand user base like film→digital→phone. H2: complete commercial loops on precise small scenes; keep monthly iteration across market/product/tech.
— END —