wjluoxiao/XB_ToolBox

What it solves

XB_ToolBox addresses the steep learning curve and high hardware requirements of ComfyUI, particularly for beginners. It solves the problem of "noodle wire" workflow clutter, the difficulty of understanding complex spatiotemporal parameters for video generation, and the frequent Out of Memory (OOM) errors encountered when running large 3D-DiT video models on consumer-grade GPUs.

How it works

The toolkit combines front-end UX enhancements with low-level memory management. It uses a "Mirror Clone Console" to centralize scattered nodes into a single dashboard and provides visual radar tools to illustrate how image and video chunks are sliced. For memory optimization, it employs "Spatiotemporal Chunk Hijacking" and dynamic VRAM offloading, which moves specific model blocks to system RAM and aggressively clears PyTorch cache to fit large models into limited VRAM.

Who it’s for

It is designed for AI beginners and creators using ComfyUI who want to simplify their workflows or run high-end video models (like WAN/LTX) on hardware that would otherwise be insufficient.

Highlights

• Extreme VRAM Optimization: Includes a Sampler Chunk Master for spatiotemporal slicing and a Model Block Swap tool to prevent OOM on 14B-22B models.
• Visual Learning Tools: Features a dual-zone radar for visualizing spatial chunks and temporal cylinder stacks.
• Unified Control: Provides Master nodes for image and video parameters to ensure dimensions and frame counts follow safe rules (e.g., the 1+8N rule).
• Workflow Organization: Offers a "Zen" Dashboard for centralized control and a Dynamic Bus node to reduce visual clutter in complex graphs.