COMPARISON
ComfyUI workflow or online? An honest comparison
The LTX-2.3 Clean Plate model is open-weight — you can absolutely run it yourself. Whether you should depends on your hardware, your volume, and how you feel about node graphs at 1 a.m.
Running it in ComfyUI
The official route: install ComfyUI-LTXVideo, download the LTX-2.3-22B base checkpoint (tens of GB; FP8 and GGUF quants exist), drop the ~330 MB Clean Plate LoRA into models/loras, and load the IC-LoRA video-to-video example workflow. Feed your clip as the reference, write a prompt describing the empty scene, render.
- Hardware: realistically a 24 GB GPU for HD; quantized variants run on less with slower, softer results
- Setup:an afternoon if you’re comfortable with ComfyUI; longer if you’re new
- Cost: free per run (your electricity), unlimited volume — the right call for heavy daily use
- Control: full — every sampler setting, LoRA strength, and resolution is yours to tune
Running it here
- Hardware: none — cloud GPUs run the full 22B model at full precision
- Setup: zero; the official prompt template and recommended settings are baked in
- Cost: 1 free clip/day, then ~13–15¢ per output second
- Best for: occasional use, laptops, and anyone who wants the result more than the pipeline
The short version
Daily volume and a 24 GB card? Set up ComfyUI — the model is free and the control is worth it. A handful of clips a month, or no GPU? The online route costs less than the coffee you’d drink during the install, and the weights are identical.
Deep dive: how the LTX-2.3 Clean Plate model works