Block the shot in Blender, render it in LTX: the Layout-to-Render LoRA

Prompting 'slow dolly left, then crane up' and hoping is still how most AI shots get their camera move. This LoRA from Lightricks takes a different approach: you set the move in a real 3D tool. It's an IC-LoRA (video-to-video, reference-conditioned) trained on LTX-2.5-22B. You give it two things. The first is a layout clip: a grey clay viewport render, or just a blocky playblast of simple shapes, from Blender, Unreal or similar. The second is a first frame that sets the art direction. The output follows the layout's camera and object positions and takes its look from that frame. For honesty about freshness: the files went up on Hugging Face in late September (last updated Oct 1), and LTX pushed it on X on Oct 15, which is when we spotted it.

The recipe from the model card: to make the first frame, take frame one of your clay animation and restyle it with any image model. Don't feed in the grey frame itself, or you get a grey video. Add a mid and a last keyframe if the look drifts. Use a 24 fps layout clip with width and height divisible by 64 (1920×1088 is the known-good size). Frame counts snap to 8×k+1, so a 72-frame clip comes out as 65, and that's expected. Write a prompt of one or two sentences that describes the finished shot, and keep the words clay, 3D and Unreal out of it.

In ComfyUI: put the LoRA in models/loras, install the ComfyUI-LTXVideo nodes (about 4.2k GitHub stars when we checked), and open the two-stage distilled workflow JSON included in the repo. Load the LTX-2.5-22B distilled weights: the transformer, the gemma4-12b-with-proj text encoder, the video and audio VAEs and the x2 latent upscaler. Set the LoRA to 1.0. Stage one renders 960×544 in 8 steps, then the latent is upscaled 2× and stage two runs 3 steps at 1920×1088, with CFG 1 and euler_ancestral. The card also shows the same job as a Python command line, which suits scripted batches.

Why we care: this fits the agent-plus-Blender workflows we keep covering, like Kevin Ngo handing a 2D piece to a second Claude session that rebuilt it in Blender. An agent can block a scene through a Blender MCP, render a quick playblast, and hand it to this LoRA for the final look. It also means a client's storyboard camera can survive into the final shot. Caveats: it's a 22B model, so check your VRAM before you plan around it (we haven't measured it). It's under the LTX-2 community licence, so read that before commercial use. And we haven't rendered with it yet.