liormamon/ltx23-bgswap-iclora-run4

BG-Swap — identity-locked background replacement for LTX-2.3

An IC-LoRA that moves a real person in a video to a new location. The person stays exactly themselves. The background comes from a single image, and the person is relit toward the new scene. One inference, no rotoscoping.

Project page: https://liormamon.com/ic-lora/ ComfyUI workflow and setup notes: https://github.com/LIORMAMONKIKI/bg-swap

Which file to use

file	what it is
run5_lora_weights_step_03000.safetensors	use this one. The released model.
run5_lora_weights_step_02000.safetensors, run5_lora_weights_step_02500.safetensors	earlier checkpoints of the same run, for comparison
lora_weights_step_03000.safetensors	run 4, person reference only, no background image. Superseded.
run6_lora_weights_step_03000.safetensors	run 6, an experiment that added single-frame training rows. Measured equal to run 5 on video. Kept for reference.

Inputs

• Person video: 1280×704, 97 frames, 25 fps
• Background image: 1280×704
• Prompt: describe the new scene and its lighting. Do not describe the person; identity comes from the video reference.
• Negative prompt: empty

Settings (official LTX single-stage distilled recipe)

Base ltx-2.3-22b-dev → ltx-2.3-22b-distilled-lora-384 at 0.5 → this LoRA at 1.0. cfg 1, euler_ancestral_cfg_pp, 8 distilled sigmas, single stage at full resolution, both reference guides at latent_downscale_factor 1. Do not use cfg above 1 or the two-stage upscale pipeline: the first makes the person flicker, the second destroys identity.

Training

Trained with the official LTX-2.3 trainer on pairs of real footage: the same person composited into another clip's background (the condition), the untouched original as the target, plus the target scene's clean plate as a second reference. 43 clips, 108 rows with lighting variants, 3000 steps, rank 32.

License

LoRA weights: MIT. Base model weights are distributed by Lightricks under their own license.