ai-fusion-video: an agent-driven platform for AI short dramas and motion comics

What ai-fusion-video sets out to do

ai-fusion-video, shown in Chinese as Rongguang, is a self-hosted platform for producing AI short dramas, motion comics and videos end to end. Rather than being a single generator, it drives the full production pipeline, script to storyboard to images to video, with agents coordinating the steps. It is built on a Java backend, a Next.js frontend and the agentscope 2.0 agent runtime, and it is deployed as a set of containers. The audience is creators and small studios who want one platform to run a short-drama workflow on their own infrastructure, wiring in the image and video models they already pay for. It is an orchestration and management layer, so it presupposes you bring model providers and are willing to run a real backend rather than use a hosted service.

A pipeline coordinated by agents

The mechanism is a task pipeline with agents at each stage and a task center that tracks runs. The changelogs describe the moving parts concretely: a pipeline whose sub-tools record their calls, an AI task center that surfaces those results, and per-task polling intervals, timeouts, cancellation and output-node selection. A significant addition is ComfyUI workflow support, covering provider configuration, workflow import, version management, validation, trial runs and publishing, so images and videos can be generated through imported ComfyUI workflows and a model can be bound to a published workflow. Project settings carry through to script and storyboard titles, and there is timezone configuration defaulting to Beijing time. The design is that of a production tool: state is tracked, runs are cancellable, and the platform coordinates external generation rather than performing it itself.

Deploying with Docker Compose

The repository contains a Java backend, a Next.js frontend and Docker Compose deployment configuration, and Compose is the recommended path. After installing Docker Engine and Docker Compose, you clone the repository and bring the stack up:

git clone https://github.com/Stonewuu/ai-fusion-video.git
docker compose up -d

On first start you follow an initialization wizard to create an administrator account, then begin using it. You can check the stack and follow logs with docker compose ps and docker compose logs -f backend, and a separate docker-compose.build.yml builds the images locally with docker compose -f docker-compose.build.yml up -d --build. The .env.example in the repository is where provider and environment configuration lives, which you fill in before pointing the platform at your models.

Where a self-hosted pipeline gets heavy

The limitations are those of running a full application. This is a Java plus Next.js stack deployed through Docker Compose, so operating it means managing containers, storage and updates, which is a real commitment compared with a hosted creative tool. It orchestrates generation rather than performing it, so you still need image and video model providers or ComfyUI workflows and whatever those cost, and the quality of the output is set by the models you connect, not by the platform. The documentation and interface are Chinese-first, with an English README provided, which is worth knowing if your team does not read Chinese. And it is a young platform whose changelogs are dominated by stability fixes, such as guarding against crashes in the task center, which is the normal texture of a fast-moving self-hosted product.

The platform versus stitching tools by hand

The alternative most creators use today is a chain of separate tools: a chat model for the script, an image model for stills, a video model for motion, glued together by hand or by ad-hoc scripts, or a commercial short-video generator that hides the pipeline. Hand-stitching gives full control and no platform to run, at the cost of managing every handoff and losing a shared record of the production. A commercial generator is simplest but closed and usually not self-hosted. ai-fusion-video sits between them: it gives you one self-hosted platform that tracks the whole pipeline and lets you bind your own models and ComfyUI workflows, in exchange for operating the stack. If you produce short dramas repeatedly and want an owned, auditable pipeline, that trade can pay off; for a one-off video it is more infrastructure than the job needs.

MIT license and release cadence

ai-fusion-video is MIT-licensed, so the code is freely reusable, and it targets a modern stack with Java 21 and Next.js 16 per its badges. The last push was on 2026-08-18, and the releases through early August 2026 show active work: ComfyUI workflow support in one version, then a run of stability fixes for the task center and client error handling in the next. Expect that pattern, capability additions followed by hardening, from a platform at this stage. Before adopting, plan the provider side first, since the platform's value depends entirely on the image and video models or ComfyUI workflows you connect, and budget for running and updating a multi-container deployment rather than treating it as a quick install.

Editorial conclusion

Adopt ai-fusion-video if you produce AI short dramas or motion comics repeatedly and want one self-hosted, agent-driven platform that tracks the whole pipeline and binds to your own models and ComfyUI workflows. Do not choose it for a single video or if you cannot run and maintain a Java plus Next.js Docker stack, and expect to supply and pay for the generation models yourself. Deploy it with docker compose up -d, complete the initialization wizard to create an admin account, and configure your providers in .env before your first run.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. Stonewuu/ai-fusion-video on GitHub