LTX 2.3 Director + Motion Workflow [Fast + Minimal + Low VRAM] [ANY T2V-I2V-V2V] — v1.0

Text, image, video-guidance and audio combinations

Generated or imported synchronized audio

DaSiWa two-stage draft → spatial upscale → refinement pipeline

Optional user LoRAs and live sampling previews

Optional body-and-hand motion transfer

Session-cached pose processing for faster repeat runs

Optional temporal 2× FPS, tiled VAE and output processing

Performance

Measured with the default inputs and base workflow settings:

RTX 4070 Ti SUPER: approximately 120 seconds

With a new motion-transfer video: approximately 400 seconds

Times vary with resolution, duration and enabled options. Reusing an unchanged motion video within the same ComfyUI session skips pose preprocessing through the session cache.

Motion-transfer warning

Motion transfer is experimental and currently not very accurate . It works best for broad body and hand movement, but should not be expected to reproduce choreography precisely. Fast motion, unusual anatomy, occlusion and facial movement can drift significantly. Source-video appearance and audio are ignored—the motion route transfers pose guidance only.

Credits

Heavily inspired by DaSiWa’s original LTX 2.3 workflow and quality settings.

Please visit and support DaSiWa / Darksidewalker on Civitai .

Models

DaSiWa LTX 2.3 — DragonLeap v4

LTX 2.3 Comfy files—VAEs and text projection

Gemma 3 12B Heretic v2 FP8

Official LTX 2.3 spatial/temporal upscalers

Union Control IC-LoRA

Cond-safe distilled LoRA

DWPose TorchScript model

Director nodes: WhatDreamsCost-ComfyUI .

Checkpoint · Hugging Face · hr16/DWPose-TorchScript-BatchSize5