I finally got around to testing the LTX Video 0.95 model in ComfyUI, and here’s how it went. First things first—I made sure my ComfyUI was up to date. The LTX integration works best with the latest version, so I hit “Update All” in the Manager just to be safe.
Dealing With Node Issues
The Video Helper nodes weren’t showing up at first, which was annoying. A quick refresh fixed that, but it’s one of those little quirks you run into sometimes. If you’re having trouble, the ComfyUI Discord usually has answers for weird issues like this.
Grabbing the Right Models
For the models, I grabbed ltx-video-2b-v0.9.5.safetensors and dropped it into the checkpoints folder.
The T5 text encoder was next—I went with the fp16 version since my GPU could handle it, but there’s an fp8 option if you’re tight on VRAM. The model files are on the LTX examples page if you need them.
T5 Text Encoder:
- Download : t5xxl_fp16.safetensors or t5xxl_fp8_e4m3fn.safetensors
First Attempt With Text-to-Video
The text-to-video workflow was straightforward. I loaded the LTX model, set a prompt like “a cat playing in snow,” and adjusted the frame count for a 4-second clip. On my 4090, it generated in under 20 seconds, which was way faster than I expected.
Where It Struggles
The motion wasn’t perfect—especially for organic shapes like animals. The cat ended up looking more like a fuzzy blob than an actual cat, but for quick experiments, it’s solid. I found that adding more descriptive terms to the prompt helped a bit.
Trying Image-to-Video
Image-to-video worked even better. I fed it a starting frame, lowered the compression value to 5, and got smoother results. The model handles transitions surprisingly well if you guide it with multiple keyframes.
Pro Tips for Better Output
One thing I noticed: detailed prompts help. Instead of just “waterfall,” specifying “a waterfall flowing rapidly with mist” gave much better motion. And if the output’s a bit rough, running it through Topaz Video AI later cleans it up nicely.
Final Thoughts
Overall, LTX Video 0.95 is a fun tool for rapid prototyping. It’s not flawless, but the speed makes up for it. If you’ve tried it, let me know—I’m curious how others are using it. The official LTX examples page has more workflows if you want to dive deeper.
