Cseti/LTX2.3-22B_IC-LoRA-Cameraman_v1

LTX-Video 2.3 22B — IC-LoRA: Cameraman v1

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

Example ComfyUI workflow

You can find a ComfyUI workflow example here: https://huggingface.co/datasets/Cseti/ComfyUI-Workflows/blob/main/ltx/2.3/ic-lora-cameraman/README.md

Example outputs

Each video shows the reference (left) and generated output (right) side by side.

How It Works

During inference you provide:

• A reference video that carries the desired camera motion
• A text prompt describing the scene to generate

The model transfers the camera behavior from the reference into the generated output. No trigger word is required.

Training Details

This IC-LoRA was trained on RunPod cloud GPUs.

Parameter	Value
Base model	LTX-Video 2.3 (22B)
Training framework	ltx-trainer (Lightricks)
Training strategy	IC-LoRA (video_to_video)
Best checkpoint	step 10,500
LoRA rank / alpha	32 / 32
Target modules	attn1, attn2 (to_k/q/v/out), ff.net.0.proj, ff.net.2
Learning rate	1e-4 (linear decay)
Mixed precision	bf16
Batch size	1 (gradient checkpointing enabled)
Training dataset	77 video pairs
Resolution buckets	768x512x57; 768x512x89; 768x512x121
First frame conditioning	0.2

Dataset

77 video pairs annotated by camera motion type, balanced to up to 15 samples per motion component. Some compound motions (e.g. zoom_in + tilt_up, orbit_cw + pan_left) are also represented.

Motion	Samples
zoom_in	15
zoom_out	15
tilt_up	15
tilt_down	9
pan_left	15
pan_right	15
orbit_cw	15
orbit_ccw	15

Usage

Requires the ltx-trainer repo and its dependencies.

uv run python -m ltx_pipelines.ic_lora \
    --distilled-checkpoint-path /path/to/ltx-2.3-22b-distilled.safetensors \
    --spatial-upsampler-path /path/to/spatial_upsampler.safetensors \
    --gemma-root /path/to/gemma \
    --lora lora_weights_step_10500.safetensors 0.8 \
    --video-conditioning /path/to/reference.mp4 1.0 \
    --prompt "Your scene description here" \
    --width 768 --height 512 --num-frames 97 \
    --output-path output.mp4

• --video-conditioning: reference video carrying the camera motion to replicate, followed by conditioning strength
• --lora: path to this LoRA followed by strength (0.7–1.0 recommended)
• No trigger word needed

Tips

• If the camera motion transfer feels too subtle, explicitly describe the desired movement in the prompt. This can strengthen the effect.

Limitations

• First experimental IC-LoRA checkpoint — results may vary
• Complex compound motions may not transfer reliably
• Only tested with I2V (image-to-video) conditioning — T2V mode is untested

License

Apache 2.0

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: