Kimi K3, Blender MCP and ComfyUI video tutorial

This tutorial follows Kimi K3 inside Hermes as it builds a Blender mountain scene with a train and camera movement. The presenter connects the AI agent through Blender MCP, supplies a reference image and asks it to create the scene. The demonstrated Hermes setup requires Blender, the MCP add-on and the Hermes skill; the presenter connects the server through Blender's sidebar panel.

The scene takes about half an hour in this demonstration. It still needs human review: some objects appear in the wrong places, including a train on the lake. The presenter supplies screenshots and asks the agent to correct the layout rather than editing Blender controls directly.

For video generation, the agent removes the scene's colors and exports a gray clay MP4 that retains the camera motion. ComfyUI then uses this animation with LTX 2.3 and the 3DREAL Strong v2 LoRA. A Krea 2 workflow with ControlNet depth creates the first-frame reference, which guides the finished video's coloration. The presenter shows a 1080p result and warns that the LoRA's trigger text must remain in the first prompt line.

The presenter describes Blender and LTX processing as local, but does not establish that Kimi K3 inference runs on the workstation. The closing section contrasts ComfyUI Cloud MCP with a separate local MCP integration, explains API JSON export and defers installation to an earlier video.