Uploaded March 2026 | Updated November 2026, 3 days ago
I have created a custom LTX 2.3 ComfyUI workflow. It uses a secret three-stage setup to fix stiff motion and broken audio.
How do you fix stiff AI video? I built a three-stage workflow that solves this entirely
. Instead of the standard two-stage setup, my workflow generates your video and audio at a low resolution first
. Then, I pass that output directly into a 2x or 4x upscaler
. This stops the camera from freezing, saves video memory, and gives you highly realistic motion
The File Setup: I made the setup incredibly simple. You just need four specific files
. You download the main 22B Dev model, the spatial upscaler, the Distilled LoRA, and the Gemma 3 text encoder
. Put them in your models folder
. Four files. Four folders. It is exactly that easy
LTX 2.3 LoRA: 5 Ways to Edit ANY Video With Text: youtube.com/watch?v=2piBFWVIUbs
LTX 2.3 In-painting Workflow: youtube.com/watch?v=lQcter3uVQA
LTX 2.3 ComfyUI Workflow: The ID-LoRA Perfect Lip Sync Fix: youtube.com/watch?v=-NPcoFF6QkU
LTX-2.3 ComfyUI vs Seedance 2.0: youtube.com/watch?v=t_y0je2dN0o&t=2s
My Camera LoRA Trick: If your prompt does not move the camera, I have a fix. I use the old LTX-2 Dolly Out camera LoRA
. But there is a major problem with this file. Running it at full power breaks your generation completely and causes visual artifacts
. That is exactly why I drop the strength down to 0.5
. This low strength protects the video while still moving the camera perfectly
The Resolution Rule: I designed this workflow for specific sizes
. For horizontal video, I use 320 by 244
. For vertical phone video, I use 244 by 320
. If I want more detail, my recommended starting point is 320 by 512
Workflow: aistudynow.com/how-i-fixed-ltx-2-3-video-artifacts-comfyui-3-stage-workflow
I have created a custom LTX 2.3 ComfyUI workflow. It uses a secret three-stage setup to fix stiff motion and broken audio.
How do you fix stiff AI video? I built a three-stage workflow that solves this entirely
. Instead of the standard two-stage setup, my workflow generates your video and audio at a low resolution first
. Then, I pass that output directly into a 2x or 4x upscaler
. This stops the camera from freezing, saves video memory, and gives you highly realistic motion
The File Setup: I made the setup incredibly simple. You just need four specific files
. You download the main 22B Dev model, the spatial upscaler, the Distilled LoRA, and the Gemma 3 text encoder
. Put them in your models folder
. Four files. Four folders. It is exactly that easy
LTX 2.3 LoRA: 5 Ways to Edit ANY Video With Text: youtube.com/watch?v=2piBFWVIUbs
LTX 2.3 In-painting Workflow: youtube.com/watch?v=lQcter3uVQA
LTX 2.3 ComfyUI Workflow: The ID-LoRA Perfect Lip Sync Fix: youtube.com/watch?v=-NPcoFF6QkU
LTX-2.3 ComfyUI vs Seedance 2.0: youtube.com/watch?v=t_y0je2dN0o&t=2s
My Camera LoRA Trick: If your prompt does not move the camera, I have a fix. I use the old LTX-2 Dolly Out camera LoRA
. But there is a major problem with this file. Running it at full power breaks your generation completely and causes visual artifacts
. That is exactly why I drop the strength down to 0.5
. This low strength protects the video while still moving the camera perfectly
The Resolution Rule: I designed this workflow for specific sizes
. For horizontal video, I use 320 by 244
. For vertical phone video, I use 244 by 320
. If I want more detail, my recommended starting point is 320 by 512
Workflow: aistudynow.com/how-i-fixed-ltx-2-3-video-artifacts-comfyui-3-stage-workflow










