You don't really direct a video model, so you impose the staging on it.

3DREAL Strong v2, a LoRA by fal for the LTX 2.3 video model, uses Blender scenes and keyframes to control AI video geometry and style.

The principle is based on a division of labor. A rough 3D scene, made of untextured gray boxes, establishes the geometry: action, trajectories, camera placement, timing, and cut points. Image models define the style, using approved references and one keyframe per shot. What remains is to convert the raw material into final pixels: this is the role of 3DREAL, a LoRA trained for the LTX 2.3 video model, of which Lovis Odin and fal are releasing a new version, Strong v2.

The model receives the 3D clip itself as dense conditioning (motion, layout, camera) plus a keyframe as an appearance anchor. Trained on sequences pairing 3D renders and photorealistic footage of the same composition, the LoRA learns to repaint surfaces without altering the geometry. Three adapters adjust the intensity of the transformation, from the most faithful to the most aggressive; Strong v2 adds shot-to-shot consistency within the same clip, synchronized audio generation, better handling of speaking characters, and a sharper image.

The demonstration is an action short film featuring a 1930s pickup truck jumping over a steam train in a desert town. The entire Blender scene—modeling, animation, cameras—was built by Claude Fable via the Blender MCP, without manual intervention on the geometry. The author emphasizes the method rather than the magic trick: three prompts, one per phase, with a human validation step between each, plus verification tools that the pipeline builds for itself (ray-traced visibility checks, fifty-percent overlay of keyframes on the gray render to track perspective drift).

The 3DREAL weights are published on Hugging Face, and a hosted endpoint on fal executes the same processing.