Open question · tool selection
What's the perfect tool stack for a fully-automated AI digital-human video pipeline?
We broke the pipeline into 6 stages and listed the 2026 candidates plus our current leanings for each. This isn't "tell us what to use" — it's putting our thinking on the table for people who've actually shipped a fully-automated avatar line. Goal: fully automated, daily-publishable, multilingual, platform-passing.
6 stages · candidates & leanings
Script / topic
Leaning: LLM batch scripting (Claude / GPT / Gemini)
Hook templates, trending-topic mining. We want event-driven, not a PPT read-aloud.
Avatar / lip-sync
Leaning: HeyGen / Synthesia (stable) · Hedra (lip-sync rated 9/10) · D-ID (cheapest API)
Kling 3.0 (anatomically consistent), DeepBrain, Colossyan (generous free tier), Elai (clone yourself). Blocker: non-English lip-sync + identity stability + no plastic look.
Voice / TTS / cloning
Leaning: Commercial: ElevenLabs; Open-source: Coqui XTTS v2 (multilingual, local)
Resemble, built-in TTS. Blocker: naturalness + emotion + alignment with lip-sync.
B-roll / footage
Leaning: Image-to-video fill (Kling / Runway / Veo)
Stock + screen capture + real footage mix. Blocker: visual consistency with the avatar.
Edit / compose / captions
Leaning: Remotion (code-driven) / CapCut · auto-captions
FFmpeg scripted batch. Blocker: pacing (a beat every 8–12s) + consistent caption style.
Orchestration / automation
Leaning: n8n to chain it all (script → TTS → avatar → edit → publish)
Homegrown Python pipeline + queue. Blocker: which stage must keep a human, which can go fully auto.
Why ask in the open
This space has too many tools changing too fast — single-tool reviews don't help. The hard part is the combination: which few pieces actually chain end-to-end, cheaply, at daily cadence. We'd rather show our reasoning and get corrected than burn money trial-and-erroring in private.
4 questions for people who've shipped
- If the goal is fully-automated, daily-publishable, multi-language, platform-passing — what would you pick for each of the 6 stages? Which combo has the best ROI?
- Lip-sync + voice emotion in 2026: is HeyGen good enough for non-English, or is Kling/Hedra + ElevenLabs better?
- Which stages MUST keep a human gate (or quality collapses), and which can safely go fully auto?
- How do you keep cost down? Per-clip SaaS billing vs open-source self-host (XTTS / local models) — where's the break-even?
Our constraints
- Primarily Chinese content; target platforms include Douyin / Xiaohongshu / WeChat Channels / YouTube Shorts
- Must support daily cadence — so per-clip cost and human intervention should be as low as possible
- No real faces, no face-swap; the avatar is an original virtual character
- We have local compute (NAS / own machines) — open-source self-host is on the table, not just SaaS
How to respond
If you've actually shipped a fully-automated digital-human line — SaaS combo or open-source self-host — come tell us your stack. We'll compile the responses into a public "digital-human pipeline selection" comparison, with attribution. This is the open selection clinic of "AI tools, the hard way."
Open question · stack updated over time · we'll fold working setups back into this page.