Creating cinematic AI videos with realistic camera movement is still difficult. Most video generation models can animate scenes, but they struggle to accurately reproduce specific camera motions from a reference video.
Movements like cinematic pans, smooth zoom-ins, orbit shots, tilts, or compound camera actions often look inconsistent or unnatural. Even when the subject looks correct, the virtual camera behavior can feel robotic or disconnected from the original reference.
 |
| cameraman ic lora showcase(transferring motion from ref video) |
Cameraman v1 aims to solve this problem by teaching LTX-Video 2.3, how to replicate camera movement directly from a reference video. Built as an In-Context LoRA (IC-LoRA) adapter, the model focuses specifically on video-to-video camera motion transfer. Instead of generating random movement, it learns cinematic motion patterns such as: zooms, pans, tilts, orbits, and combined motion sequences.
The model is based on LTX-Video 2.3 (22Billion) and was trained using the ltx-trainer framework by Lightricks with an IC-LoRA video-to-video strategy.
The dataset was carefully collected by camera motion type and balanced across different movement categories. It includes both simple and compound camera actions such as-zoom_in+ tilt_up, orbit_cw +pan_left. The training also used multiple resolution buckets- 768x512×57,768x512x89,768x512×121. Additionally, first-frame conditioning was set to 0.2 to help maintain scene consistency while transferring motion behavior.
Installation
1. Make sure you have the ComfyUI installation. Update it if using the older version from the Manager.
2. Download and setup the basic Ltx2.3 (I2V) models and workflow.
3. Now. download Ltx2.3 IC LoRA Cameraman v1 (LTX2.3-22B_IC-LoRA-Cameraman_v1_10500.safetensors ) from its hugging face repository. Then, save this into ComfyUI/models/loras folder.
4. Restart ComfYUI to take effect.
Workflow
1. Download the workflow (LTX-2.3-IC-Lora-CameramanV1.json) from our hugging face repository.
2. Drag and drop into comfy.
3. Load your reference image and reference video into load image/video node.
4. Load other models (Ltx2.3, vae, text encoders) into their respective nodes. Load Ic lora cameraman v1 into lora model loader only node.
5. Put prompts into prompt box.
This is the intial variant trained on lesser dataset so you may get bad weird results. Make sure you donot put the more descriptive prompts related camera movements as it can contradict and may produce bad results. Model can easily sense and replicate the motion from your reference video.
6. Hit run to start generation.
Cameraman v1 highlights an important shift in AI video generation: controlling how a scene is filmed is becoming just as important as generating the scene itself.
Most current models focus heavily on character consistency or visual quality, but believable camera movement is what often separates amateur-looking AI videos from cinematic ones. What makes this LoRA interesting is its specialization.
Rather than trying to solve every video problem at once, it focuses narrowly on motion language and that targeted approach appears to deliver stronger results. For creators working with LTX-Video pipelines, this could become a valuable tool for achieving more professional and controlled cinematography without massive retraining costs.
