ComfyUI out of memory: the Dynamic VRAM fix that works

If your character reel dies at 80% with CUDA error: out of memory while Task Manager shows free VRAM, the card is probably not the problem. Since March 2026 ComfyUI has loaded models through Dynamic VRAM, and on 14 October 2026 the three most-reported crashes under it are still open on the tracker.

This guide is the order that costs the least time: read the error string, run one diagnostic launch flag, pin one package if the string says so, and only then decide you need a bigger GPU. Every flag and version below is checked against Comfy's own documentation or issue pages, and anything I could only read in a third-party guide is labelled as such.

Why free VRAM and an out-of-memory error can both be true

The old loader estimated how much memory a model needed and reserved it up front. Dynamic VRAM, announced by the maintainer on 1 March 2026, copies the parts of a model the GPU needs at the moment it needs them, so a model can be larger than your VRAM or your RAM. The stated goals were less RAM use, faster first loads and LoRA changes, and fewer out-of-memory stops.

It is on automatically on Nvidia GPUs under Windows and Linux. Comfy's startup flags page lists it as automatic on Nvidia and on ROCm 7.14 or newer, and the maintainer said in the announcement that WSL may never be supported. That is why a card with 12 GB can run a 33B-parameter video model at all, as covered in the MiniMax H3 review.

The cost is a new failure surface. Streaming weights through pinned host memory, reading them from disk and mapping them into the GPU can fail in places the old loader never went, and those failures often surface as a generic out-of-memory message. The announcement thread also warned that higher VRAM use is normal and that Task Manager may not show lower RAM, because weights stay cached in RAM.

Read the error string first

Different crashes have different first moves. Match the exact text in your console, not the symptom. The table pairs each string with the issue where it is tracked and what reporters or maintainers tried first.

The fix order that works, one change at a time

Change one thing per launch. A troubleshooting guide that tracks these reports makes the same point: if two flags change together, a success teaches you nothing.

Flags that no longer do what old guides say

Most advice on the web was written before March. Two flags are the usual traps.

docs.comfy.orgOpen the page

One more trap, from the troubleshooting guide above: --normalvram is no longer a valid argument and can crash the desktop app. And do not stack --disable-smart-memory, --reserve-vram and --cache-none hoping one helps; the guide advises against combining them, and you will not learn which one mattered.

What is still broken on 14 October 2026

A fix you can trust needs a merged change, and these are not there yet. Nothing below is a prediction; it is what the pages said when I read them.

The pattern: Comfy's maintainers are working on it, none of these three issues has a fix you can install today, and the disable flag is the one workaround reporters agree on, with the caveat that it does not rescue everyone.

When it really is the card

Sometimes the answer is capacity. A 33B video model on a 12 GB card runs on offloading and goodwill, and long clips can fail at the decode step. Shorten the clip, lower the resolution, or drop the batch to one before you buy anything. If a smaller job reliably moves the failure point and nvidia-smi shows memory pinned at the ceiling, rent a bigger GPU for the long renders instead of fighting the loader. Our Wan 2.2 guide and the open video model ranking say which models are kind to small cards.

If ComfyUI will not start or connect after a flag change, a mistyped flag is the usual cause; the connection help page walks the checks, and a missing node pack after pruning custom nodes is covered in the missing nodes guide.

Keep a character feed running while you debug

The expensive part of an OOM crash for a creator is not the minute lost, it is the reel that dies at the last frame and the reference photos you re-upload to retry. Musetta runs on your own ComfyUI, so these launch flags are yours to set, and it keeps your characters' reference sets in one place so a retry after a flag change is one request instead of a rebuild.

Musetta studioDemo data

The short version of the whole guide: read the string, disable Dynamic VRAM once to learn whether the loader or the card is at fault, pin the one package the string names, and keep notes. When Comfy lands a fix for the issues above, remove your workaround and take the speed back.