Rexgent: One Idea In, a Whole Mini-Drama Out

COVER

Track Winner, Global AI Hackathon Series with Qwen Cloud

Project Introduction

AI short dramas are everywhere now — flooding TikTok, Reels, and Shorts. But making one still means chaining together separate tools for script, images, voice, video, and editing, with a human being the glue holding it all together. Rexgent replaces the human glue. It's an autonomous AI showrunner: give it a one-line premise or your own screenplay, pick one of 22 visual styles, and it turns it into a finished, voiced short drama — while optimizing quality, budget, and token usage the entire way.

Project Value

Here's the insight that drives everything: writing the script is no longer the hard part. Everything after the script is. Characters change faces between shots. Scenes misread the intended action. Later scenes contradict earlier ones. Voices fall out of sync. And video costs spiral out of control because every regeneration burns real money.

Rexgent attacks each failure mode directly. It writes the script, scores its own quality_, and rewrites weak drafts — so the story is strong before a single dollar goes to video. It locks each character's face and outfit into reference plates, gives every character a distinct voice, and maps how they all connect. An AI Director storyboards every scene with real camera language, while a narrative memory stops the story from contradicting itself. During rendering, faces and dialogue go to HappyHorse for native lip sync while silent scenery goes to Wan, and every clip is scored for drift and flagged rather than silently shipped. Finally, Rexgent assembles the episode with captions and music, ready to download. Every token and dollar is tracked and predicted _before you spend it. Run it fully automatic, or stay in control at every stage. One idea in, a whole drama out.

Solution

Rexgent is built on a LangGraph pipeline of agents powered by 17 Qwen Cloud models. You set a title, choose one of 22 visual styles, define a budget and episode count, then type a premise or import your own script in English or Chinese. Full Auto runs the entire production; manual mode lets you review each stage, with a live crew dock and cost meter showing every agent at work in real time.

The writing agent develops the premise into cliffhanger episodes, and you can rewrite any draft, check its quality score, or run a deeper story analysis before committing. Rexgent then extracts every character and, from one reference photo, generates identity plates that lock each face and exact outfit. Each character gets a previewable voice, and a relationship graph shows how they connect. Next, an AI Director breaks every scene into shots with shot size, camera angle, lens, and blocking — enforcing the 180-degree rule so the cuts actually work. A Neo4j narrative memory keeps later scenes from contradicting earlier ones, and a top-down camera plan shows exactly what each shot sees. During rendering, each shot goes to the video model built for it, and every clip carries forward from the previous frame and is scored for drift, flagged for review if a face slips. Finally, a crew assembles the episode, placing captions on the shot that speaks each line and stitching the clips together.

Highlights

Rexgent runs entirely on Alibaba Cloud. Qwen Cloud powers 17 AI models, Alibaba Cloud OSS stores every media asset from character plates to final episodes, and the whole platform is deployed on Alibaba Cloud ECS using Docker Compose. The guiding principle: no model is expected to do everything. Each stage is routed to the most suitable model based on capability _and cost_.

Qwen-Max handles the creative work — story development, screenwriting, shot direction — where generation quality matters most. Qwen-Plus does the analytical work: narrative judging, plot-gap detection, prompt refinement, relationship extraction. Qwen-Flash handles deterministic, high-volume work — scene parsing, character extraction, wardrobe, set dressing, titles — at a fraction of the cost. For vision, Qwen3-VL-Plus scores clips for continuity and narrates each clip's final frame so the next shot opens from ground truth, while Qwen-VL-Max reads uploaded reference photos. For image, Wan2.6-T2I generates the production-bible reference plates and Qwen-Image-Edit-Max creates consistent costume variations from a locked character identity. For video, HappyHorse 1.1 handles characters and native speech, Wan 2.7 renders scenery and silent continuation, and HappyHorse-Video-Edit fixes a flagged take. For audio, Qwen3-TTS provides preset, designed, and cloned voices, while Fun-ASR and Qwen3-AS
R-Flash deliver word-level timing and hallucinated-speech detection. By reserving premium models for creative tasks and routing everything else to the cheapest capable model, Rexgent cuts token cost without sacrificing quality.

Alibaba Cloud Products Used

About the Team

Rui Zhe (Rexton) is a final-year Bachelor of Computer Science (Artificial Intelligence) undergraduate at the University of Malaya, a nine-time hackathon winner, and currently a Data Scientist Intern, with a passion for turning complex data into actionable insights and building end-to-end AI products. Rexgent — the name blends "Rexton" + "agent" — was his first video-generation project and his first time building on Alibaba Cloud and Qwen, done entirely solo on a budget of just $40 in Qwen Cloud credit. That constraint became the lesson: "AI slop in video is expensive. A bad script costs a few cents to rewrite, but a bad video clip costs real money every time you regenerate it." With no room to reroll, he was forced to plan everything before rendering — score the script first, lock the faces first, plan every shot first — and that planning discipline became the very core of Rexgent. Having text, image, video, voice, and vision models all under one platform, he notes, is what let one person build a full production pipeline that would normally need a whole team of tools.

Explore the project: https://www.youtube.com/watch?v=xq4kIzE0r0w

This project was developed as part of the Global AI Hackathon Series with Qwen Cloud.

Community

Rexgent: One Idea In, a Whole Mini-Drama Out

Project Introduction

Project Value

Solution

Highlights

Alibaba Cloud Products Used

About the Team

Read previous post:

Alibaba Cloud Project Hub

You may also like

Comments

Alibaba Cloud Project Hub

Related Products

Token Plan

Qwen

Alibaba Cloud Model Studio

QwenWork