LTX 2.3 + IC-LoRA Cameraman: Image to Video

LTX 2.3 image-to-video with a Cameraman IC-LoRA that copies camera motion from a reference clip.

Upload a starting image, drop in a reference video with the camera move you want (a pan, a dolly, a slow push-in), write what your scene should be, and run it. Your image animates with the borrowed camera language. Audio is generated in the same pass, so if your prompt includes spoken lines, the model attempts to lip-sync.

How do you use IC-LoRA Cameraman with LTX 2.3?

Upload your starting image. Load a reference video with the camera move you want. Write a positive prompt describing your scene and any dialogue. The Cameraman IC-LoRA reads the reference and applies that motion to your image, then a second pass upsamples the result. Hit Run.

Starting image This is frame one of your video. The model animates forward from here. A clean, sharp image holds up best. Faces and small details degrade fast if the source is soft.

Reference video This is where the camera motion comes from. The IC-LoRA looks at the movement in this clip (pans, dollies, push-ins, handheld sway) and copies it onto your scene. The subject in the reference does not have to match yours. Only the camera language transfers.

Positive prompt Describe what you want to see: subject, action, look, setting. If your subject speaks, put the line in quotes. LTX 2.3 generates audio in the same pass and will try to sync the mouth. Keep prompts visual and grounded. Avoid stacking mood words like "moody" or "cinematic"; the model amplifies them and they get hard to escape.

Negative prompt Tells the model what to avoid. Default is "pc game, console game, video game, cartoon, childish, ugly" to keep things in a realistic register. Edit if you want a different look. Leave it as-is for photoreal output.

IC-LoRA strength (default: 1.0) How hard the reference's camera motion is imposed on your scene. 1.0 follows the reference closely. Drop to 0.6 or 0.7 if the motion is too aggressive for your shot. Lower still if you want a hint of the move rather than a copy.

bypass_i2v toggle Leave False for image-to-video. Flip to True for text-to-video. The model then generates the first frame from your prompt with no input image.

Resolution and length Defaults to 1280x720, 97 frames at 25fps (about 4 seconds). Width and height are exposed as their own nodes. Longer clips eat more VRAM and time. Stick near 720p for the cleanest results from this distilled checkpoint.

The defaults are tuned. For a first run, change only the image, the reference video, and the prompt.

What is LTX 2.3 + Cameraman IC-LoRA good for?

Locking in a specific camera move you have seen somewhere else. Drop in a reference clip with the shot you want, like a Steadicam push-in, a whip pan, or a slow crane lift, and the IC-LoRA transfers that motion to your generated scene without you needing to describe it in words.

Useful when prompt-only camera direction is not getting you the move. Phrases like "slow dolly in" or "handheld pan" are ambiguous to most video models. A reference clip is not.

Good for previz, look-dev shots, mood reels, music videos, and any case where the camera move matters as much as the content. If you have a director's reference and want to test your own scene in that language, this is the workflow.

Not the right tool for clean text-to-video from scratch with no camera intent. A regular LTX 2.3 t2v workflow is faster. Also not ideal for multi-shot edits in one generation. One image, one camera move, one continuous take is the sweet spot.

FAQ

What is the Cameraman IC-LoRA for LTX 2.3? It is an In-Context LoRA that reads camera motion from a reference video and applies it to a new generation. Instead of training a custom motion module or fishing for the right prompt words, you hand it a clip with the move you want and the model copies the camera language onto your scene.

Does the reference video need to match my subject? No. The IC-LoRA only borrows the camera motion: pans, dolly moves, zooms, handheld feel. Your subject, setting, and style come from your prompt and starting image. A skateboard clip can guide the camera for a beauty shot. A drone clip can guide a portrait.

Why does my LTX 2.3 output have audio? LTX 2.3 22B is multimodal. It generates video and audio in the same pass. If your prompt includes dialogue in quotes, the model attempts to lip-sync. To remove audio, mute the output player or strip the track from the file after download.

How long is the generated video? Default is 97 frames at 25fps, about 4 seconds. You can push longer by raising the frame count, but VRAM and run time scale with it. The two-pass setup (base generation plus latent upsampler) is tuned for the 4-second range.

Can I use this for text-to-video instead? Yes. Flip the bypass_i2v toggle to True and skip the starting image. The model generates the first frame from your prompt and the IC-LoRA still imposes camera motion from your reference clip. The starting image is the only thing that gets bypassed.