Most of the time wasted on a fresh RunPod pod isn't GPU time. It's the first twenty minutes figuring out why a download failed or why the disk looks full when it shouldn't be. None of these are hard problems once you know they exist. They're just not obvious until they bite you.
df is lying to you
Run df -h on a network volume and you'll get a number that looks like your available space. It isn't. On a network volume, df reports space across the whole cluster the volume lives on, not what's actually free for your pod. You can have plenty of room and still see a number that makes you think you're about to fill the disk.
The number that actually matters is your own usage:
du -sh /workspace
That walks your workspace directory and tells you how much you've actually put there. It's slower than df because it's doing real work instead of reading cached cluster stats, but it's the only number worth trusting when you're deciding whether you have room for another model pack.
If you're setting this up for the first time, the RunPod template page has the baseline config this is built around, so you're not guessing at paths or mount points.
Why the volume is worth the setup
The reason to bother with a network volume instead of the pod's local disk is persistence. Models on a network volume survive pod restarts. You download a checkpoint, a LoRA pack, a video model — once — and it's there the next time you spin up a pod, even if that pod is a different instance entirely.
Without a network volume, every fresh pod means re-downloading everything from scratch. For a single checkpoint that's annoying. For a full set of models across image, video, and upscale workflows, that's a real chunk of your session gone before you've generated anything. Pay for the volume, download once, and the next session starts with everything already in place.
This matters more as your workflow list grows. If you're running image generation, character work, and video upscaling off the same pod, you're accumulating a lot of model weight over time. A volume that persists is the difference between a five-minute cold start and a forty-minute one.
The Civitai token trap
Here's one that looks like a network problem but isn't. You try to download a model from Civitai with a direct link, no token attached, and the download appears to succeed. File shows up, right size in the ballpark, no error thrown.
Then you load it in ComfyUI and it fails, or worse, loads garbage. What actually happened: without an API token, Civitai serves you a login page instead of the file. That login page gets saved to disk with the filename you expected, and depending on your download tool, it can look enough like a real file that you don't notice until it breaks something downstream.
The fix is just making sure your token is attached to every Civitai download, not just the first one. It's easy to set it up correctly once, then lose it on a later command, especially if you're scripting downloads for several model packs in one session.
Quick gut check before trusting a download: open the file size. A saved login page is tiny — kilobytes, not gigabytes. If a checkpoint you expected to be multiple gigabytes shows up at a few KB, that's not a corrupted download, that's a login page wearing a model's filename.
Picking the right card
The other trap is overpaying for a GPU tier you don't need, or underpaying and hitting a wall mid-session.
For the Gmanski workflow set — image generation, character work, standard video — a 48 GB card like an A6000 or L40S handles everything. You don't need to reach for anything bigger by default.
The exception is LTX-based long video work. An 80–96 GB card doesn't change what's possible there, but it's noticeably faster, which matters if you're processing long clips and billing by the hour.
If you're only running stills and shorter video, don't pay for the bigger card. If long-form LTX upscaling is a regular part of your pipeline, the faster tier pays for itself in reduced pod time.
Takeaway
Before you touch a model: check space with du -sh /workspace, not df. Confirm your Civitai token is attached before any download, and verify file size on anything that matters. Then size the GPU to the workflow, not the other way around — a 48 GB card covers the standard set, and only LTX work needs the bigger tier.
