Start with a true story from the spring of 2026 — it will save you a few thousand yuan. Jimeng AI’s premium annual card cost ¥2,599/year with 15,000 monthly credits and a first-year 60% discount; two months later it became ¥3,099/year with 6,160 monthly credits and the discount gone — the cost of the same 15-second video rose 7× overnight. On Xianyu, the secondhand marketplace, “legacy discount annual accounts” were resold for ¥5,000 to ¥20,000 and jokingly called “the best-performing fund of the year.”
That story compresses the entire lesson of tool selection: you are picking tools, platforms are picking prices, and the latter can change at any time. So this part’s shopping lists follow one principle throughout — lists are consumables; principles are assets. Part 1’s pipeline diagram (script → storyboard → images → motion+voice → edit) now becomes concrete tools, prices, and two standing rules.
The five evaluation dimensions: consistency outranks everything
Score every AI video tool on the same table; don’t get carried away by demo reels on official sites:
- Character consistency — the heaviest weight, bar none. Viewers forgive a blurry frame; they will never forgive the protagonist growing a new face in shot ten. The human brain is far more sensitive to face consistency than to scenery — it’s the shared hard problem of every 2026 model, and the dividing line between tools.
- Cost per second — ignore the sticker price, do the division: monthly fee ÷ monthly credits ÷ per-second burn. The 2026 market runs ¥0.2–1.0 per second — a 5× spread.
- Speed and queue — 1–5 minutes per clip is normal; free tiers queue for hours at peak. Throughput is a pipeline property, not a single-point property.
- Chinese comprehension — prompts, captions, dialect lip-sync. Domestic models win this axis by default; overseas models rendering Chinese poster text is basically a lottery.
- Commercial license and watermark — required reading before monetization. Whether free-tier output can be used commercially and whether it carries a watermark varies wildly across platforms — and as Part 1 noted, revenue-share contracts will audit your material sources.
Axes one and five are veto items; the other three determine the value ranking.
The four-layer review: who plays which position
Four pipeline layers, the mid-2026 candidate list, one-line positioning:
Text-to-image (character and scene assets)
- Jimeng: top-tier Chinese comprehension, out-of-the-box character lock, wired into the Jianying editing ecosystem — the default first pick for Chinese creators.
- Midjourney ($10–60/mo): strongest artistry and stylization, ideal for character design sheets and key frames, but needs an access workaround.
- SDXL local: free, no content review, pixel-level control via ControlNet/IP-Adapter — the ceiling for the technical route, and the steepest learning curve.
Image-to-video (the money layer) — SuperCLUE’s April 2026 Chinese image-to-video leaderboard top three: PixVerse V6, Seedance 2.0, Vidu Q3 Pro. Pick by positioning:
- Kling 3.0: the value king at roughly ¥0.43/second (annual top-tier card math); strongest motion performance, 4K/60fps, native audio-video output since 2.6; free users get 128 monthly credits (~14 seconds).
- Vidu Q3: the pioneer of character consistency, highest anime-style affinity, 16-second clips with native sound; off-peak generation is cheaper — the natural choice for comic-drama creators.
- Hailuo 2.3: unmatched action dynamics and fight scenes (~¥0.93/second), but silent by default and still strengthening multi-shot narrative — first pick for action, detour for dialogue.
- Jimeng/Seedance 2.0: a generational lead in audio-video sync, strongest Chinese, fastest generation; the price is the most violent pricing volatility (see the opening story).
Voice / TTS
- Doubao TTS: multi-character voices plus emotion control, same ecosystem as your script tool — first choice.
- Jianying built-in voices: free, one-click, lip-sync matching — enough to start.
- Xunfei (¥30–200/mo): dialects and broadcast tones; Moyin Workshop (¥99–599/mo): custom voice cloning — the key tool for character IP-ification.
Editing
- Jianying Pro: free, auto-captions, smart cutout, and direct timeline import of Jimeng assets — the de facto standard of the Chinese pipeline.
- CapCut: the overseas edition, for international distribution. Premiere/AE: finishing and effects, high learning cost — reconsider at the ¥2,000 tier.
The script-layer LLM sits above all four layers: Doubao for dialogue, Kimi for long web-novel adaptation, DeepSeek/Gemini for storyboard breakdown (exactly the combination in production teams’ public workflows).
Three shopping lists
- ¥0/month (prove the pipeline): Jimeng’s daily free quota + Kling’s 128 monthly credits + Doubao + the Jianying suite. Enough for 4–8 episodes a month. The goal of this stage isn’t money — it’s proving you can walk all five stages reliably.
- ¥500/month (the minimum viable toolchain): Kling gold membership (¥58–66) as the base + Doubao TTS or Xunfei + Vidu off-peak generation for action shots + Jianying Pro. Supports 15–20 episodes a month — serialization cadence. Of Part 1’s five revenue streams, novel referrals and custom production already work at this tier.
- ¥2,000/month (serial-production tier): Midjourney + Jimeng premium + multi-platform credits + Moyin voice cloning + Premiere. Don’t enter this tier before your first revenue-share payout — tool upgrades should be a consequence of revenue, not a prerequisite.
Two standing rules
Rule one: pay monthly, never yearly. The Jimeng annual-card lesson is industry norm, not anecdote — credit shrinkage, discount removal, model rebranding; platforms hold a hundred legal ways to reprice. The 20% a yearly card saves you corresponds to a 700% loss risk when rules change. Every tool: monthly. All credits: spent within the month.
Rule two: keep a backup on every layer. Primary Kling means backup Vidu; primary Jimeng means backup Kling; primary Doubao TTS means backup Jianying. Platforms don’t notify you before changing rules, and one day of pipeline downtime costs more than a backup membership. Also write the gacha rate into your budget: the 2026 floor is a 5–10% failure rate per generation, so reserve 2–3 rerolls per asset — that’s not waste, it’s the cost of yield.
Action items, then Part 3
- Assemble the ¥0 toolchain: register Jimeng, Kling, Doubao, and Jianying; use each one’s free daily quota to generate a 5-second test clip; record generation time and your subjective quality score — your first measured data point.
- Do one division: pick two image-to-video tools and compute true per-second cost via “monthly fee ÷ monthly credits ÷ per-second burn,” then compare against this part’s ¥0.2–1.0/s range. This habit will inoculate you against every future pricing pitch.
- Read one set of terms: find your primary tool’s commercial-license and watermark clauses and answer three questions: can free-tier output be monetized? Do revenue-share platforms accept it? Can the watermark be removed?
Part 3 enters the heaviest stage of the pipeline: script and storyboard. How the viral short-drama structural formula (the golden 3 seconds, a hook every 15 seconds, the closing cliffhanger) translates into an LLM prompt workflow, how to turn the storyboard table into a reusable template, and why “topic > script > production” is the only correct priority order in this track.