LTX 2.3 Video Pipeline (Pose Control + Face/Hand Detailer)

This is a highly optimized, production-ready video-to-Video workflow for LTX-Video 2.3. It perfectly syncs dance motions using IC-LoRA Union Control while preventing any facial or hand distortion through a dedicated post-processing detailer pipeline.

Key Features

• Base Engine: LTX-Video 2.3

• Motion Control: Integrated IC-LoRA-Union-Control for flawless pose tracking.

• Quality Preservation: Advanced Face & Hand Detailer pipeline to fix distortion during intense movements.

• Hardware Friendly: Tailored for mid-tier GPUs (e.g., RTX 4060) using FP8 Scaled Transformer.

Nodes description & caution

Please strictly follow these parameter rules to avoid artifact burning or system errors:

• Photo and Pose Reference Input
  Once you have prepared your photos, insert the image extracted from the first frame into "First Image" and the image extracted from the last frame into "Last Image."
  

• Estimate the Pose
  it used the DWPose estimate and Depth, which is fast and does not require the installation of a separate large model. Depth offers better performance, so try setting it to Depth to extract poses and create a video!

  • Additional files required : depth_anything_vit
    

• 1st Sampler Strength Adjustment

  • For dynamic videos, you need to adjust the strength level slightly. Lowering the strength reduces distortion of the human body, but it may decrease the accuracy of the pose.

• Face and Hand Detailer

  • Denoise : You can draw faces and hands more accurately by adjusting the denoise value. A maximum value of 0.2 is recommended.

  • Additional files required : face_yolo8, hand_yolo8

• Upscaling

  • Upscaling is not yet perfect. If you have a good sigma value, please adjust it or replace the scale model to make adjustments.

📦 Required Models

• Z Image Turbo

• ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors