LTX2.3-ICEdit-Insight

LTX2.3-ICEdit-Insight

<table> <tr> <td align="center"><img src="./assets/effects/output_004.webp" alt="Video restoration preview" width="420"/></td> <td align="center"><img src="./assets/effects/视频高清对比效果.webp" alt="Video HD enhancement preview" width="420"/></td> </tr> <tr> <td align="center"><img src="./assets/effects/去水印对比效果.webp" alt="Watermark removal preview" width="420"/></td> <td align="center"><img src="./assets/effects/去字幕对比效果.webp" alt="Subtitle removal preview" width="420"/></td> </tr> </table>

LTX2.3-ICEdit-Insight is a task-aware video restoration and editing model family developed by JoyFox Lab, built on top of the LTX-2.3 DiT-based audio-video foundation model.

This release focuses on four practical video editing directions:

• Video Restoration: degradation recovery, compression cleanup, blur and noise reduction, and damaged detail restoration.
• Video HD Enhancement: super-resolution, detail reconstruction, texture sharpening, and perceptual quality improvement.
• Watermark Removal: logo cleanup, semi-transparent overlay removal, and occlusion-aware background reconstruction.
• Subtitle Removal: hard subtitle removal, caption cleanup, text overlay removal, and temporally stable inpainting.

Unlike conventional frame-level enhancement pipelines, this model family operates as a generative video restoration system in latent video space. It is designed to preserve global structure, camera motion, object identity, and temporal consistency while reconstructing missing or degraded visual content.

Project links: GitHub project | JoyFox on Hugging Face | Paper (Research Square) | DOI

📄 Paper

• Title: LTX-Insight: Unified Video Restoration and Semantic Editing via Task-Aware Adaptation and Temporal Consistency
• Authors: Fan Tang, Siyuan Li
• Venue: Research Square
• Posted date: May 25, 2026
• DOI: 10.21203/rs.3.rs-9775063/v1
• Paper URL: https://www.researchsquare.com/article/rs-9775063/v1

This model release corresponds to the paper's unified video post-processing framework built around three high-level settings: video super-resolution/enhancement, occlusion removal and repair, and instruction-driven semantic editing.

In the currently released inference package, those ideas are exposed through four practical routes:

• Video Restoration
• Video HD Enhancement
• Watermark Removal
• Subtitle Removal

The paper introduces three core components:

• TADA (Task-Aware Decoupled Adapter) for separating structure restoration from texture/semantic editing under a shared backbone.
• FGLC (Flow-Guided Latent Consistency) for explicit temporal consistency modeling in latent space using optical-flow alignment and cross-frame attention bias.
• PMTR (Progressive Multi-Task Routing) for more stable joint multi-task training through curriculum scheduling and collaborative routing.

📦 Model Files

🎬 Showcase

<table> <tr> <td align="center"><b>Video Restoration</b></td> <td align="center"><b>Video HD Enhancement</b></td> </tr> <tr> <td align="center"><img src="./assets/effects/output_004.webp" alt="Video restoration preview" width="600"/></td> <td align="center"><img src="./assets/effects/视频高清对比效果.webp" alt="Video HD enhancement preview" width="600"/></td> </tr> <tr> <td align="center"><b>Watermark Removal</b></td> <td align="center"><b>Subtitle Removal</b></td> </tr> <tr> <td align="center"><img src="./assets/effects/去水印对比效果.webp" alt="Watermark removal preview" width="600"/></td> <td align="center"><img src="./assets/effects/去字幕对比效果.webp" alt="Subtitle removal preview" width="600"/></td> </tr> </table> <br/> <table> <tr> <td align="center"><b>Video Restoration</b></td> <td align="center"><b>Video HD Enhancement</b></td> </tr> <tr> <td align="center"><img src="./assets/effects/视频修复对比效果2.webp" alt="Video restoration preview 2" width="600"/></td> <td align="center"><img src="./assets/effects/视频高清对比效果2.webp" alt="Video HD enhancement preview 2" width="600"/></td> </tr> <tr> <td align="center"><b>Watermark Removal</b></td> <td align="center"><b>Subtitle Removal</b></td> </tr> <tr> <td align="center"><img src="./assets/effects/去水印对比效果2.webp" alt="Watermark removal preview 2" width="600"/></td> <td align="center"><img src="./assets/effects/去字幕对比效果2.webp" alt="Subtitle removal preview 2" width="600"/></td> </tr> </table>

🚀 Script Usage

Run all scripts from the project root.

bash run_restoration.sh
bash run_hd.sh
bash run_hd.sh /path/to/input.mp4
bash run_watermark_rm.sh
bash run_watermark_rm.sh /path/to/input.mp4
bash run_subtitle_rm.sh
bash run_subtitle_rm.sh /path/to/input.mp4

💻 Command Examples

Video Restoration

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
python run_pipeline.py \
  --mode restoration \
  --video ./inputs/input_480p.mp4 \
  --prompt "Convert the video to ultra-high-definition quality while removing artifacts and rebuilding high-frequency details." \
  --output ./outputs/output_restoration.mp4 \
  --height 1184 --width 704 --num-frames 97 \
  --fps 24.0 --seed 42 \
  --sigma-profile workflow \
  --streaming-prefetch-count 2 \
  --model-checkpoint ./models/checkpoints/ltx-2.3-edit-insight-dev-fp8.safetensors \
  --lora ./models/loras/ltx2.3-train/ltx2.3-video-restoration-general.safetensors

Video HD Enhancement

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
python run_pipeline.py \
  --mode hd \
  --video ./inputs/input_480p.mp4 \
  --prompt "Convert the video to ultra-high-definition quality, significantly improving clarity, fine detail richness, texture fidelity, and overall perceptual sharpness." \
  --output ./outputs/output_hd.mp4 \
  --height 1184 --width 704 --num-frames 97 \
  --fps 24.0 --seed 42 \
  --sigma-profile workflow \
  --streaming-prefetch-count 2 \
  --model-checkpoint ./models/checkpoints/ltx-2.3-edit-insight-dev-fp8.safetensors \
  --lora ./models/loras/ltx2.3-train/ltx2.3-ic-video-upscale-general.safetensors

Watermark Removal

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
python run_pipeline.py \
  --mode watermark_rm \
  --video ./inputs/input_480p.mp4 \
  --prompt "Remove short-video platform watermarks and related occlusions from the video, restoring a clean, clear, and natural original image." \
  --output ./outputs/output_watermark_rm.mp4 \
  --height 1184 --width 704 --num-frames 97 \
  --fps 24.0 --seed 1546 \
  --sigma-profile workflow \
  --streaming-prefetch-count 2 \
  --model-checkpoint ./models/checkpoints/ltx-2.3-edit-insight-dev-fp8.safetensors \
  --lora ./models/loras/ltx2.3-train/ltx2.3-ic-watermark-remove-general.safetensors

Subtitle Removal

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
python run_pipeline.py \
  --mode subtitle_rm \
  --video ./inputs/input_480p.mp4 \
  --prompt "Remove subtitles, captions, and related text occlusions from the video, restoring a clean and natural underlying image." \
  --output ./outputs/output_subtitle_rm.mp4 \
  --height 1184 --width 704 --num-frames 97 \
  --fps 24.0 --seed 42 \
  --sigma-profile workflow \
  --streaming-prefetch-count 2 \
  --model-checkpoint ./models/checkpoints/ltx-2.3-edit-insight-dev-fp8.safetensors \
  --lora ./models/loras/ltx2.3-train/ltx2.3-ic-subtitles-remove-general.safetensors

✨ Key Improvements

Task-Aware IC-Edit Framework

We introduce a task-aware IC-Edit training framework for LTX-2.3, where each restoration direction is optimized with dedicated instruction conditioning and task-specific IC-LoRA adapters.

The model is trained not only to improve visual quality, but also to understand the editing goal behind different restoration tasks, including watermark removal, subtitle cleanup, damaged region recovery, and high-definition enhancement.

LTX-2.3 DiT Backbone Adaptation

The model family is built on the LTX-2.3 foundation architecture, a diffusion-transformer video model designed for high-fidelity image-to-video and video generation workflows.

Our adaptation targets video restoration by improving:

• latent-space editability
• instruction-following behavior
• frame-to-frame stability
• high-frequency detail recovery
• local reconstruction around degraded or occluded regions

Spatiotemporal Consistency Optimization

Video restoration requires more than strong single-frame quality. We optimize temporal consistency so that restored areas remain stable across adjacent frames.

This reduces common artifacts such as:

• flickering textures
• unstable reconstructed backgrounds
• inconsistent watermark removal
• subtitle ghosting
• frame-wise color shift
• detail popping during motion

Degradation-Aware Training Curriculum

The training curriculum covers realistic video defects including:

• compression artifacts
• motion blur
• sensor noise
• low-bitrate video
• text overlays
• hard subtitles
• semi-transparent watermarks
• platform logos
• local occlusions
• low-resolution inputs

This improves generalization across short videos, social-media clips, mobile footage, downloaded videos, and compressed production material.

Occlusion-Aware Reconstruction

For watermark and subtitle removal, the model is optimized to reconstruct the hidden visual content behind occluded regions.

Instead of smearing or blurring the target area, it uses surrounding spatial context and temporal cues to infer plausible background structure, object boundaries, lighting, and texture continuity.

Frequency-Enhanced HD Restoration

For HD enhancement, the model improves perceptual sharpness and fine visual detail through frequency-aware restoration training.

This is especially helpful for recovering:

• hair strands
• fabric texture
• skin detail
• product edges
• background patterns
• typography-like fine structures
• natural image clarity

🧠 Inference Notes

• Single-stage inference is recommended for most editing tasks.
• Two-stage refinement can improve visual polish but may weaken task-specific LoRA constraints.
• Watermark and subtitle removal perform best when the occlusion area is stable and not excessively large.
• HD enhancement quality depends on input resolution, motion complexity, and compression level.
• Higher output resolution improves detail but requires more VRAM.
• For strong-motion videos, conservative denoising settings are recommended to preserve temporal structure.
• Frame count should follow the 8k + 1 rule.
• Output height and width should be multiples of 32 in single-stage inference.

📚 Citation

If this model family is useful in your research or application workflow, please cite:

@article{tang2026ltxinsight,
  title   = {LTX-Insight: Unified Video Restoration and Semantic Editing via Task-Aware Adaptation and Temporal Consistency},
  author  = {Tang, Fan and Li, Siyuan},
  journal = {Research Square},
  year    = {2026},
  month   = {05},
  doi     = {10.21203/rs.3.rs-9775063/v1},
  url     = {https://www.researchsquare.com/article/rs-9775063/v1}
}

🏗️ Training

This model family was trained and optimized by JoyFox Lab (Chengdu Xuanhu Technology Co., Ltd.).

The training pipeline includes:

• task-aware video restoration data construction
• degradation synthesis and curriculum training
• IC-LoRA specialization for four editing directions
• temporal consistency regularization
• occlusion-aware reconstruction training
• high-frequency perceptual enhancement
• instruction-guided video editing optimization

📬 Contact

For research collaboration, commercial licensing, or workflow integration, contact:

• [email protected]

📜 License

Licensed under Apache 2.0.

Please also review the license terms of the upstream LTX-2.3 base model when using or redistributing derivative checkpoints.