8 Days, 20 USD, One CLI: Building an Open-Source AI Manhua-Video App with Claude Code & GLM-4.7

Core question answered in one line:
A backend-only engineer with zero mobile experience can ship an end-to-end “prompt-to-manhua-video” Android app in eight calendar days and spend only twenty dollars by letting a CLI coding agent write Flutter code while a cheap but powerful LLM plans every creative step.

1. Why Another AI-Video Tool? The Mobile Gap

Core question this section answers:
If web-based manhua-video makers already exist, why bother building a mobile-native one?

• Every existing product the author tried was desktop-web only, asking users to upload reference images, drag storyboards, and wait for cloud queues—none of which feels natural on a bus ride.
• During a Wuhan AI meet-up, thirty attendees repeated the same pain: “I want to create on my phone while commuting.”
• The author’s earlier two blog posts on AI manhua drew tens of thousands of reads and dozens of private “when will there be an app?” messages.
• Personal motivation: the author’s wife joked “you must be crazy” when hearing the idea—proving her wrong became an eight-day sprint target.

Author reflection:

I realised the competition wasn’t other apps; it was the empty time people spend scrolling. If creation is easier than consumption, they’ll create.

2. Success Metric: What Does “Done” Look Like on Day 8?

Core question:
Under money, time, and skill ceilings, what exact artefact counts as victory?

Checklist used every evening:

• [x] 1-sentence prompt accepted
• [x] 8-scene script auto-written
• [x] 1 protagonist, 1 side character, 7 scenes configurable
• [x] Three-view character sheet generated once and reused
• [x] Each scene: 1 key image + 4 s video clip
• [x] FFmpeg concatenates clips into final MP4
• [x] MIT license repo public on GitHub

3. Tech Choices in 10 Minutes: Flutter + GLM-4.7 + ReAct Loop

Core question:
Which stack can be believed in without proof-of-concept time?

1. Coding Agent
   Claude Code (CLI) showed 95 % compile-success on Flutter snippets in earlier toy tests—no GUI, no config hell, just yes to every suggestion.

2. UI Framework
   Flutter: one code-base → Android APK, hot-reload < 1 s, default Material3 theme already “pretty enough”.

3. LLM Brain
   GLM-4.7:

   • Chinese & English scripting equally fluent
   • JSON instruction format stable across 50 prompt iterations
   • Year-end promo: 100 M tokens for 20 USD top-up

4. Media API

   • Image: Gemini Pro Vision endpoint (author had ready-made doc)
   • Video: Veo 2.0 beta (author had ready-made doc)

5. Control Pattern
   ReAct loop:
   User text → LLM thinks → JSON tool call → App executes → result fed back → LLM next action … until “action: finish”.

Author reflection:

I picked Flutter over React-Native simply because Claude Code hallucinates fewer imports in Dart. That’s not engineering elegance—that’s deadline survival.

4. Data-Flow Architecture: One Picture, No Black Box

Core question:
How does a plain sentence become a watchable 30-second manhua video?

flowchart LR
    A[User types prompt] -->|1| B[GLM-4.7 planner]
    B -->|JSON tool| C[Flutter app]
    C --> D{Router}
    D -->|write_script| E[Local YAML]
    D -->|draw_char| F[Gemini image]
    D -->|draw_scene| G[Gemini image]
    D -->|gen_video| H[Veo API]
    F & G & H -->|URL| I[Feedback to GLM]
    I --> B
    B -->|action:finish| J[FFmpeg concat]
    J --> K[final.mp4]

Consistency trick:

• Same seed and negative_prompt block across all calls.
• Character LoRA trained on three-view sheet, weight 0.8.
• Video calls use the key image as first frame reference.

5. Day-by-Day Log: From 0 Lines to APK in the Store Folder

Core question:
What does the actual grind look like when the clock never stops?

Worst moment: 2 a.m. on D5, Veo threw “invalid aspect ratio” 7 times—docs said 16:9 instead of 9:16. One word, two nights of sleep lost.

6. Walk-through: 32-Second “Strawberry Cake” Manhua

Core question:
Can a reader reproduce an entire clip right now with nothing but the repo?

Step 0 Input

User types:
“a pink-haired girl baking a strawberry cake, cute vibe”

Step 1 Script (GLM-4.7)

{
  "title": "Berry Sweet",
  "scenes": 8,
  "hook": "Her heart beats louder than the oven timer.",
  "protagonist": { "name": "Berry", "trait": "pink bob, strawberry badge" }
}

Time: 3.2 s | Tokens: 1.1 k

Step 2 Three-View Sheet

Prompt snippet:
「pink bob, strawberry badge on apron, three-view sheet, seed128475, negative: blurry, extra limbs」
Output: 1024×1024 PNG, face IoU 0.87

Image source: author generation

Step 3 Key Images ×7

Example scene-3 prompt:
「close-up, Berry placing strawberry on cream mountain, window light, seed128475」
Generation: 6 s, 1024×1792

Step 4 Video Clips ×7

Request:

{ "image": "<key_image>", "duration": 4, "motion": "subtle head tilt, cream swirl" }

Median latency 28 s, 4 s@30 fps MP4 returned.

Step 5 Concatenate

No re-encode:

ffmpeg -f concat -safe 0 -i list.txt -c copy final.mp4

Size: 32.4 MB, 720×1280, 32 s.

Play-test screenshot:

Image source: author generation

7. Character Consistency Deep Dive: Seed, LoRA, Reference Frame

Core question:
How is face-drift kept under 10 % without manual touch-up?

1. Global Seed

   • Three-view, key images, and video first-frame all reuse seed=128475.
   • Identical negative_prompt block removes random limb chaos.

2. Lightweight LoRA

   • Train 20 steps, rank 16, on three-view sheet → 3.7 MB file.
   • Inference weight 0.8: keeps face, allows style flexibility.

3. Reference-Frame Video

   • Veo accepts first-frame image; feed resized 512×512 crop of face.
   • Motion descriptor limited to subtle or slow to avoid warping.

Numbers from 20-run ablation:

• Face IoU ≥ 0.85: 18 / 20
• Human rating ≥ 4 / 5: 9 / 10 clips

8. Budget Autopsy: Where Did the 20 USD Go?

Core question:
Is “cheap” a marketing line or a repeatable fact?

Promo detail: 20 USD top-up during campaign gave 100 M tokens. At list price (0.06 USD / 1 k) the same run costs ~78 USD—still under the 100 USD ceiling.

9. Repo Tour & Local Build in 5 Minutes

Core question:
How can a reader clone and see her own manhua video tonight?

git clone https://github.com/<user>/man-dao.git
cd man-dao
cp config.yaml.example config.yaml
# fill GLM_KEY, GEMINI_KEY, VEO_KEY
flutter pub get
flutter run --release

Key folders:

• lib/react_loop.dart – ReAct parser, 180 lines
• scripts/seed_lock.py – enforces same seed across APIs
• assets/lora/berry_rank16.safetensors – 3.7 MB character weights

First successful compile:

Image source: Unsplash

10. Lessons Learnt & Road-map

Core question:
If the author started tomorrow, what would he skip, double-down on, or never do again?

Lessons

1. Read docs to the pixel—aspect ratio typo cost 14 hours.
2. Free tiers are great until day 7; always have a second provider URL ready.
3. Version-control the prompt—rolling back a 200-line system prompt by Ctrl-Z is not fun.

Next milestones

• Mandarin TTS with CosyVoice + lip-sync (already tested, PR pending)
• In-app sharing to mini-program (backendless, QR-code only)
• Community LoRA market so users can swap protagonists in one click

Action Checklist / Implementation Steps

1. Install Flutter 3.16 and Android Studio Hedgehog.
2. Clone repo, fill config.yaml with API keys.
3. Run flutter doctor → all ticks green.
4. Execute flutter run --release on a physical phone (camera permission needed).
5. Type a one-sentence story idea → wait 8 min → receive 30 s manhua video.
6. Train your own LoRA: put 10 three-view images under lora/train_data/ and run scripts/lora_train.py.
7. Commit, push, and tweet the repo—maintainer will merge useful PRs within 48 h.

One-page Overview

• Scope: Backend-only engineer, zero mobile exp, 8 days, 20 USD.
• Stack: Claude Code CLI → Flutter → GLM-4.7 planner → Gemini img → Veo video → FFmpeg concat.
• Loop: ReAct pattern keeps LLM in charge, app just calls tools.
• Consistency: Global seed + 3.7 MB LoRA + reference frame = ≤ 10 % face drift.
• Deliverable: 60 MB APK, open-source, MIT license, GitHub live now.

FAQ

1. Q: Can I switch to React Native?
   A: RN branch stub exists but Claude Code generates more reliable Dart; feel free to PR.

2. Q: What happens when free Veo quota dries up?
   A: Swap base_url in video_api.dart to Runway or Pika; interface identical.

3. Q: Is 20 USD a long-term realistic cost?
   A: At list price the same run costs ~78 USD; still below 100 USD cap.

4. Q: Commercial use allowed?
   A: MIT license, do as you wish; don’t upload copyrighted faces to LoRA trainer.

5. Q: iOS version?
   A: Flutter code is cross-platform; you need Apple Dev account (99 USD) and video export compliance description.

6. Q: Why English voice-over?
   A: MVP skipped language lock; Mandarin TTS PR is under review.

7. Q: Eight-day crunch healthy?
   A: Averaged 5 hrs/day, no all-nighters; double the timeline if you want weekends.