Cseti/LTX2.3-22B_IC-LoRA-Cameraman_v2

LTX-Video 2.3 22B — IC-LoRA: Cameraman v2

A fine-tuned In-Context LoRA (IC-LoRA) adapter for LTX-Video 2.3 (22B), trained to replicate camera movements from a reference video.

This is v2 of the Cameraman IC-LoRA with a larger and more diverse dataset.

Example outputs

Each video shows the reference (camera-motion input) and the generated output.

Prompt

point-of-view of a spaceship flying above an asteroid while chasing an X-Wing spaceship, firing green laser beams

Prompt

woman walks on a rainy street. the camera zooms out and upwards showing the city around

Prompt

girl is sitting on the bed while the camera pans left

Prompt

woman walking with blowing magical halo behind her head

Prompt

woman casting magic orbs with her hands

Prompt

woman in black armor stands still and starts to salute

Prompt

woman sitting on the bed holding a smoking gun

Prompt

ballerina walks slowly

Prompt

woman is walking across the room

Prompt

woman standing in a narrow storage room

Prompt

""

Prompt

""

Usage (ComfyUI)

I tested this lora only in ComfyUI. An example workflow is here: https://huggingface.co/datasets/Cseti/ComfyUI-Workflows/blob/main/ltx/2.3/ic-lora-cameraman-v2/README.md

How it works:

• Load LTX2.3-22B_IC-LoRA-Cameraman_v2_14000.safetensors as the LoRA.
• Provide a reference video carrying the camera motion you want to replicate.
• Provide a starting image. This is optional. The model works both in T2V or I2V mode.
• Provide a text prompt describing the scene to generate.
• No trigger word is needed.

Tips

• Resolution: based on my testing, the higher the resolution, the more closely the reference camera motion is followed. I wouldn't go below 960x512 for the first pass.
• Image (conditioning) strength: use an image strength of 0.5 or 0.7 for more motion.
• The prompt matters a lot — it strongly affects the camera movement. If the output doesn't follow the reference camera motion, you can try:
  • leaving the prompt empty (in some cases this works best),
  • a different seed,
  • describing the camera motion explicitly, at least at a high level.

Training Details

This IC-LoRA was trained on RunPod cloud GPUs (NVIDIA RTX PRO 6000 Blackwell, 96 GB).

Parameter	Value
Base model	LTX-Video 2.3 (22B)
Training framework	ltx-trainer (Lightricks)
Training strategy	IC-LoRA (video_to_video)
Released checkpoint	step 14,000
LoRA rank / alpha	64 / 64
Target modules	attn1, attn2 (to_k/q/v/out), ff.net.0.proj, ff.net.2
Optimizer	ProdigyPlusScheduleFree (auto-LR, prodigy_steps 1000)
Scheduler	constant (required by schedule-free)
Mixed precision	bf16
Batch size	1 (gradient checkpointing enabled)
Training dataset	343 video pairs (+ 23 held-out for validation loss)
Resolution buckets	768x512x{57,89,113,121} @ 24fps
First frame conditioning	0.3

Dataset

366 curated reference/target pairs (343 train / 23 held-out validation, 0 overlap). The set covers single-axis motions as well as many compound multi-axis combinations (e.g. pan_left + tilt_up + roll_ccw, dolly_in + truck_left + pedestal_down).

Motion-component frequency across the training set (a pair can contribute to several components):

Component	Count
pan_right	93
pan_left	90
dolly_in	83
roll_cw	79
truck_right	79
roll_ccw	77
tilt_up	75
zoom_in	69
truck_left	68
tilt_down	62
dolly_out	61
zoom_out	52
pedestal_down	50
static	28
pedestal_up	25

Of the 343 training pairs, 93 are single-axis and 250 are compound (multi-axis).

Limitations

• Complex compound motions may not transfer reliably

License

This LoRA is created as part of a personal project for research purposes only and is not intended for commercial use.

Support

Producing and sharing this kind of open-source work requires renting cloud GPUs, which gets expensive quickly. If you find it useful and would like me to keep contributing, your support is very much appreciated: