Video generation (LTX2.3) became unusable/mega-slow after days of use

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

Very first run on same workflow, prompt, ect should have consistent run time..

### Actual Behavior

Hello,

This one is a weird mysterious one, and I have no exceptions but I'm making this issue to put in your radar, maybe lots of people having same issue.

#### Problem
- Video generation became progressively slower day by day. I would restart comfyUI, then run my workflow, and notice it getting worse day by day.
- I had the exact same set up, using same workflow, no knew nodes ect, simply generating videos for days. I was using basic LTX2.3 I2V workflow, and the video generation was progressively getting slower per day. It was like if I was maxed on RAM but I wasnt.
- I would restart my computer, I could generate 1 or 2 vids, then infinity sluggishness

- Eventually (few days) I would restart computer but couldnt even gen 1 vid, just doing 1 step on low resolution would take 20 minutes per step (i should generate a 10s vid in 10 minutes or less).

- I did not have 100s of apps and 100s of tabs open, my RAM would peak at ~90% during video generation

I reinstalled today and it works again (shrug).

#### Version
- ComfyUI Version sorta v0.26.0 (I git pulled at hash 04879a8113961cbc4e2ff20e9feeb737ba703f51, which is few days before v0.26.0
- nvidia 5080 (16 VRAM)
- 64 GB system RAM
- startup script: .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-api-nodes --disable-auto-launch --disable-dynamic-vram --preview-method taesd --use-sage-attention --front-end-version Comfy-Org/ComfyUI_frontend@1.39.19
- generating a 1024x1024 vid, 241 frames, (low res at 512x512)

### Steps to Reproduce

???

### Debug Logs

```powershell
It is a silent error (the worst kind). Hangs at 0/8 steps. 20 minutes later I'll see 1/8
```

### Other

_No response_

Start at the ComfyUI/main.py entry point and reproduce the reported LTX2.3 I2V slowdown with the listed startup flags, workflow dimensions, and hardware. Compare the first few runs with later runs while collecting useful profiling or memory logs; the issue provides no deterministic steps, named tests, or code location, so done is not yet clearly defined.

Written by the indexing model from the issue text.