You need to agree to share your contact information to access this model

By clicking "Agree and Access" you acknowledge the Privacy Policy and consent to receive offers and updates. You can unsubscribe at any time.

LTX-2.3 22B IC-LoRA Beard Removal

This is a Beard Removal IC-LoRA trained on top of LTX-2.3-22B, which removes facial hair (beard, mustache, and stubble) from a person in a video while preserving their identity, expression, motion, lighting, and the surrounding scene.

It is based on the LTX-2.3 foundation model.

Model Files

ltx-2.3-22b-ic-lora-instant-shave-0.9.safetensors

A single rank-64 IC-LoRA adapter checkpoint (step 500). This is the recommended and only shipped checkpoint.

Model Details

Base Model: LTX-2.3-22B Video
Training Type: IC-LoRA
Control Type: Reference video conditioning (a clip of a bearded subject drives a clean-shaven re-render)
Reference Downscale Factor: 1 (the reference is provided at 1× the output resolution)
Pipeline details: No special pre/post-processing; standard IC-LoRA video-to-video inference.

Intended Use & Out-of-Scope

Intended use: Removing beards, mustaches, and stubble from short video clips of people, driven by a reference video, while keeping the subject's identity and motion intact. Performs best at 960×544 (landscape) or 544×960 (portrait), 25 fps, for clips of roughly 33–121 frames.

Out of scope: Not designed to grow or add facial hair, restyle hair, or alter other attributes. Quality degrades at resolutions far from the training buckets and on very long clips, where identity can drift.

Control Signal Requirements

Control signal type: Reference video (the original footage of the bearded subject).
Expected input: A single video clip.
Preprocessing: None required beyond matching the output resolution and frame rate; 25 fps is recommended.
Alignment: The reference should match the intended output in frame count, FPS, resolution, and aspect ratio.
Mask support: Not supported.

How It Works

The adapter conditions on a reference video and re-renders it with the facial hair removed, holding the subject's motion, expression, and the scene composition fixed while replacing the bearded region with smooth, clean-shaven skin. Prepend the trigger word REMOVEBEARD to the prompt and describe the clean-shaven subject and scene.

Usage

🔌 ComfyUI

Copy the LoRA weights into models/loras.
Load the LTX-2.3-22B base model and add ltx-2.3-22b-ic-lora-instant-shave-0.9.safetensors as the LoRA.
Start at strength 1.0 and adjust to taste.
Use the IC-LoRA video-to-video workflow from the LTX-2 ComfyUI repository, which wires the correct reference-conditioning nodes. Connect your source clip as the reference video and prepend REMOVEBEARD to the prompt.

Recommended Settings

LoRA strength / weight: 1.0
Inference steps: 30
Guidance scale: 4.0 (with spatiotemporal guidance / STG enabled)
Resolution & frames: Trained at 960×544 and 544×960, 49 frames @ 25 fps; generalizes well up to ~121 frames (≈5 s).
Prompting: Always start the prompt with the trigger word REMOVEBEARD, then describe the clean-shaven subject and scene ("completely smooth and clean-shaven, bare skin, no beard, no stubble, no facial hair"). Recommended negative prompt: beard, mustache, facial hair, stubble, worst quality, inconsistent motion, blurry, jittery, distorted.

References

Code: GitHub Repository
IC-LoRA docs: docs.ltx.video

Tips & Troubleshooting

Identity drift on long clips: Keep generations near the trained window (≤ ~121 frames) and at the training resolution; longer or off-resolution renders can lose the subject's identity.
Residual stubble or shadow where the beard was: Strengthen the negative prompt (beard, stubble, facial hair) and keep the trigger word at the very start of the prompt.
Best alignment: Match the reference clip's FPS to 25 and its aspect ratio to a landscape (960×544) or portrait (544×960) bucket.

Dataset

The model was trained using a proprietary dataset of paired bearded / clean-shaven video clips.

Training

Technique: IC-LoRA (rank 64, alpha 64) on the DiT transformer
Hyperparameters: bf16 mixed precision, AdamW optimizer, learning rate 4.2e-4, linear scheduler
Steps: 500 (recommended checkpoint: step 500)
Infrastructure: LTX-2 Community Trainer