JGRFW/comfyui-AICG3D

What it solves

ComfyUI users commonly hit node name collisions when installing multiple MiniMax-related plugins, which can overwrite nodes, break workflows, or crash the interface. This project solves that by enforcing a unique namespace for every core node. It also removes the need to glue together many fragmented plugins: a single package covers script splitting, rendering, upscaling, and final video assembly, letting creators produce very long, locally generated videos with predictable and stable behavior.

How it works

The plugin isolates its core nodes under an AICG3D_H3_ prefix, so other MiniMax forks installed in the same environment cannot interfere with it. It includes an intelligent script splitter that breaks long prompts into logical segments and maps them onto the video timeline, plus optimized sampling strategies, latent-space upscaling, and SelfLift enhancement. Users can reference images, audio, and video from a built-in resource library by clicking cards; these references flow through a chain of video-segment nodes that automatically inherit or override assets. Several H3 task modes (full reference, first/last frame, text-to-video, etc.) change the generation behavior accordingly. The workflow supports external samplers and DMAD acceleration, and uses FFmpeg for the final splice of the infinite-length video.

Who it’s for

Professional AI video creators, short-drama producers, and ComfyUI users who need to generate locally, without length limits, while keeping a stable and conflict-free environment. It suits both newcomers following the provided 13.x workflows and advanced users who want to build their own segment chains.

Highlights

• Zero-collision namespace: every core node uses an AICG3D_H3_ prefix, preventing conflicts with other MiniMax plugins.
• Full pipeline in one package: from prompt splitting and sampling to latent upscaling and final video composition.
• Infinite-length generation: local sequential segments, tail-frame continuation, and automatic asset inheritance across a segment chain.
• Rich integration: resource library with click-to-insert references, multiple H3 modes, external sampler inputs, and DMAD support.
• Safe loading: avoids unsafe deserialization and uses weights_only for non-SafeTensors models.
• Easy to install and discover: available via Comfy Registry search and auto-publish through GitHub Actions.