Lightricks / LTX-2.3-22b-IC-LoRA-Clean-Plate

LTX-2.3 Clean Plate IC-LoRA

An IC-LoRA for LTX-2.3-22B that removes people and other dynamic subjects from a source video and reconstructs the empty background, producing a clean "clean plate" of the same scene.

It is based on the LTX-2.3 foundation model.

Model Files

ltx-2.3-22b-ic-lora-clean-plate-1.0.safetensors

Model Details

• Base Model: LTX-2.3-22B Video
• Training Type: IC-LoRA (video-to-video)
• Control Type: Reference video — the original clip containing the subjects to remove
• Reference Downscale Factor: 1
• Audio: Not trained for audio generation (audio skipped during preprocessing)

Intended Use & Out-of-Scope

Intended use: Producing clean background plates for VFX and compositing by removing people, pedestrians, and vehicles from static or moving-camera footage while preserving the static background (architecture, ground markings, foliage, street furniture). It also removes edge/foreground occluders (e.g. a cyclist's arms, handlebars, or a bike) when enough surrounding background is visible.

Out of scope: Subjects that dominate and occlude most of the frame — for example first-person legs/torso filling the center of a POV shot — where too little background remains to reconstruct. Use dedicated mask-based inpainting for those cases.

Control Signal Requirements

• Control signal type: Reference video (RGB) — the original clip with the subjects present.
• Expected input: A video clip with spatial dimensions divisible by 32 and a frame count of 8k+1, supplied as the video-to-video reference / first-frame conditioning.
• Preprocessing: None beyond a standard resize to a 32-divisible resolution and an 8k+1 frame count. The stock workflow encodes audio, so ensure the clip has an audio track (add a silent one if the source has none).
• Alignment: The output is frame-aligned to the reference — same length and same framing.
• Mask support: None. This is a full-frame model that removes dynamic subjects globally; it does not take a user-supplied region mask.

How It Works

The model is trained video-to-video: the reference clip's latent is the conditioning signal and the target is the people-free clean plate. At inference you feed the original clip as the reference, and the LoRA regenerates it with dynamic subjects removed while keeping the static background intact. There is no trigger word — conditioning comes from the reference latent, and the text prompt is used to describe the desired empty scene (and to name specific objects that should not appear).

Usage

ComfyUI

1. Copy the LoRA weights into models/loras .
2. Load the LTX-2.3-22B base model and add ltx-2.3-clean-plate.safetensors as the LoRA.
3. Start at strength 1.0 and adjust to taste.
4. Use an IC-LoRA (video-to-video) workflow from the LTX-2 ComfyUI repository , which already wires the reference-video conditioning nodes. Feed the original clip as the reference and write a prompt describing the empty scene.

Recommended Settings

• LoRA strength / weight: 1.0
• Resolution & frames: Trained at 1024×576 and 576×1024 , 49 frames @ 25 fps. Also validated at HD 1920×1088 / 1088×1920 . Keep spatial dimensions divisible by 32 and frame counts at 8k+1.
• Prompting: Describe the desired empty scene rather than issuing a "remove X" command — the LoRA was trained on captions that describe the clean result. Put subjects that must not appear in the negative prompt.
• Name specific occluders explicitly: When a particular object must go (e.g. a bicycle), name it in both the positive ("...no bicycle or bike anywhere in the frame") and the negative prompt. In testing this was the difference between the bike/handlebars being fully removed vs. left behind.

Example positive prompt:

An empty clean plate of the exact same location: identical background, environment, structures and lighting as the source video, with no people, no humans, no figures, and no body parts such as arms, hands or legs anywhere in the frame. Static photorealistic footage, natural light, high detail.

Example negative prompt:

person, people, human, figure, limb, arm, hand, leg, body part, backpack, shadow of a person, leftover people, leftover figures, ghosting, semi-transparent remnants, blurry, soft, distorted, inconsistent motion, jitter, worst quality, cartoon, video game, ugly

References

• Code: GitHub Repository
• ComfyUI: ComfyUI-LTXVideo

Tips & Troubleshooting

• Composition matters most. Removal works best when the subject is a partial or edge occluder with visible background around it. It struggles when the subject fills the center of the frame, because little real background remains to reconstruct.
• Persisting object? Add it by name to both the positive (as "no <object> ") and the negative prompt.
• Large first-person body parts (POV legs/torso dominating the frame) are out of scope — use mask-based inpainting instead.
• Audio-encode error on the stock workflow? Ensure the input clip has an audio track; add a silent track if the source has none.
• Respect the format constraints: resolution divisible by 32 and frame count at 8k+1.

Dataset

26 paired clips. Each pair is a reference clip (a scene containing people and/or vehicles) and its clean-plate target (the same scene with the dynamic subjects removed), spanning city streets, offices, airport concourses, and outdoor/POV scenes. Mixed landscape and portrait orientation, 49 frames @ 25 fps. Captions describe the empty target scene.

Training

• Technique: Video-to-video IC-LoRA trained from scratch on LTX-2.3-22B, using reference-latent conditioning ( reference_latents_dir ) with first_frame_conditioning_p = 0.1 .
• Hyperparameters: rank 32, alpha 32, dropout 0.0, target modules attn1/attn2 {to_q, to_k, to_v, to_out.0} + ff.net.0.proj + ff.net.2, AdamW, learning rate 2e-4 (linear decay to 0), batch size 1, gradient accumulation 1, bf16 mixed precision, gradient checkpointing on.
• Steps: 3000.
• Resolution buckets: 1024×576×49 and 576×1024×49.
• Infrastructure: LTX-2 Community Trainer (single H100 80GB, ~224 min, peak ~55 GB VRAM).

License

See the LTX-2-community-license for full terms.

Acknowledgments

• Base model by Lightricks
• Training infrastructure: LTX-2 Community Trainer