Bunkley’s LTX 2.3 Director + Motion Workflow — Fast + Minimal
A compact two-stage workflow built around LTX Director 2.0 and DaSiWa’s quality pipeline.
Director acts as the workflow’s multimodal front end, allowing T2V, I2V, V2V-style conditioning, or virtually any combination of text, images, video guidance and audio. Use unlimited image keyframes, global/local prompts, generated or imported audio, audio inpainting and optional pose-motion guidance—all from one timeline.
Note: Workflow has a good preview node as well, so you can cancel incorrect generations in between, also gguf and sage attention support is present
Features
Unlimited Director keyframes and prompt segments
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
Director nodes: WhatDreamsCost-ComfyUI.