Tel Aviv · Paris · Remote
Creative technologist · Creative AI Specialist

Cinema-grade visuals at the speed of real-time.

What 3D · Video · GenAI · AR Since 2017
Scroll

My works.

Tools.

Publications.

I'm a creative technologist working in real-time 3D and generative AI. My practice combines Unreal Engine, MetaHumans, Vicon motion capture, Blender, Octane, and ComfyUI with image and video generation models.

Selected work spans film, fashion, music, and exhibition; shown internationally and featured in editorial press. I also teach on YouTube and Instagram.

Open to commissions, collaborations, and R&D partnerships.

Get in touch.

Based between Tel Aviv and Paris. Specializing in 3D animation, modeling, rendering, immersive video, GenAI, and AR.

← All works
Vibe Coding 2026

3D → Stylized · LTX-2.3 IC-LoRA

A rough 3D pass in, a fully stylized animated sequence out — with the character staying on-model across every shot. A video-to-video pipeline built on LTX-2.3, pairing IC-LoRA structural guidance with keyframe reference images.

The source is a blocky, untextured 3D/game-engine pass carrying nothing but staging and motion. LTX-2.3 re-renders it as hand-painted comic illustration while two systems hold the piece together: an IC-LoRA union guide conditions generation on a control map derived from the source video, and a designed keyframe stamps in identity and wardrobe so the character reads as the same character shot to shot.

The guide path runs DWPose, DepthCrafter and Canny preprocessors into a single selector with seven modes — pose, depth or canny alone, or any blend of the three — chosen per shot depending on what needs locking: pose for body clarity, depth for the canyon's volume, canny for edge detail, blends where one alone drifts. All three are normalized to the input video's aspect before blending, which turned out to be as much an artistic fix as a technical one; DepthCrafter normalizes to a different aspect, and the resulting stretch was quietly breaking the character's proportions.

The workflow itself was assembled programmatically — two of my graphs merged by script and structurally validated in one pass, then hardened across a run of runtime fixes and carried into production graphs for both single-GPU and 8-GPU. A snowbound canyon was chosen deliberately as the test: snow and rock punish temporal consistency, and tight corridors push depth and occlusion hard.

Role
Pipeline design · Workflow engineering · Lookdev · Direction
Stack
LTX-2.3 (IA2V) · IC-LoRA · ComfyUI · DWPose · DepthCrafter · Canny · Python
Output
102-second comparison reel at 2688×1536, cut as a workflow breakdown — stylized output against the 3D source, with the keyframe reference and live control signal inset
Note
Personal fan-art and R&D project. The showcase character is used for personal creative experimentation and remains the property of its rights holders.
3D → Stylized · LTX-2.3 IC-LoRA
3D → Stylized · LTX-2.3 IC-LoRA
3D → Stylized · LTX-2.3 IC-LoRA
3D → Stylized · LTX-2.3 IC-LoRA
3D → Stylized · LTX-2.3 IC-LoRA