Add a "Clean VRAM" option to the SeedVR2 Video Upscaler node

[ai-capybara]

Hi, and thanks for your great work on SeedVR2!

I’m using SeedVR2 for upscaling still images rather than videos. My current workflow consists of three steps:

1. Generate an image with KSampler
2. Upscale it with SeedVR2
3. Refine the result using KSampler again (with the same model as in step 1)

The issue I’m running into is related to VRAM usage. I’m using Flux.2 kein 9B or Z-Image for generation and the SeedVR2 7B models for upscaling, both of which cannot fit into my 32 GB VRAM simultaneously.

To work around this, I inserted a Clean VRAM node before the "SeedVR2 Video Upscaler (v2.5.24)" node. This helps: if I start from step 1, the VRAM is cleared properly and upscaling proceeds at normal speed.
Without cleaning, the SeedVR2 node becomes extremely slow (sometimes running indefinitely until I kill ComfyUI).

However, this workaround fails in a specific case:

• If I re-run the workflow but only change the seed in the "SeedVR2 Video Upscaler (v2.5.24)" node, ComfyUI resumes execution from that node onward.
• The preceding Clean VRAM node is skipped, leaving the previous model from step 3 loaded into memory.
• As a result, the SeedVR2 model cannot fit into VRAM, and the process hangs again.

One workaround is to add another Clean VRAM node after step 3, but this forces a re-load of the generation model when I return to step 1 which slows down iteration significantly.

Feature request:
Could you please add a "Clean VRAM before loading" option to the "SeedVR2 Video Upscaler (v2.5.24)" node?
This would ensure that, regardless of the workflow entry point, the node always frees VRAM before attempting to load the SeedVR2 model.

This feature would make complex image workflows much smoother, especially for users with limited VRAM.