Skip to main content

CUDA out of memory (ComfyUI / Stable Diffusion) — how to fix

A "CUDA out of memory" error means the model needed more GPU VRAM than the card has. It is almost always a VRAM ceiling, not a bug in your workflow — the same graph runs fine on a bigger GPU.

Also seen as: "torch.cuda.OutOfMemoryError", "CUDA OOM", "GPU out of memory"

Direct answer

How do I fix "GPU Memory Exhausted"?

A "CUDA out of memory" error means the model needed more GPU VRAM than the card has. It is almost always a VRAM ceiling, not a bug in your workflow — the same graph runs fine on a bigger GPU.

Key facts

  • iThe model + activations exceed your GPU VRAM (e.g. an fp16 14B video model wanting ~28GB on a 24GB card).
  • iResolution or batch size too high — VRAM scales with pixel count and batch.
  • iA second model (upscaler, ControlNet, VAE) loaded alongside the main checkpoint.
  • iFragmentation after many generations without a flush.

Next step

  1. 1Lower resolution first (1024×1024 → 832×1216 → 768×768) and set batch size to 1 — the cheapest fix.
  2. 2Use an fp8 / GGUF (Q5/Q4) quantized checkpoint — roughly halves weight VRAM at a small quality cost.
  3. 3Add --lowvram / --medvram (or the WanVideoWrapper block-swap / offload nodes) to trade speed for VRAM.
  4. 4If it still OOMs after that, the model genuinely needs a bigger card — move to a 48GB (A6000) or 80GB (A100) GPU.

Not a fit when

  • If it OOMs at 512×512 batch 1, quantization won't save you — the base model is too big for the card.
  • Consumer 24GB cards (4090 / L4) can't hold fp16 14B video models no matter the settings.

Cheapest cloud GPU that fits

If the model genuinely needs more VRAM, the same workflow runs unchanged on a bigger GPU. Cheapest-first:

GPUVRAMPriceAction
L424 GB$0.51/hrDeploy
A600048 GB$0.64/hrDeploy
A10080 GB$1.81/hrDeploy
H10080 GB$4.28/hrDeploy

Prices include a 30% service fee. Billed per minute while running.

Not sure this workflow will run before you rent a GPU?

Paste your ComfyUI workflow JSON into the free checker — it flags missing nodes, model requirements, and the VRAM/GPU you actually need. No signup.

Check my workflow freeSee GPU pricing

Other ComfyUI errors

Fix

Model Not Found

Fix

Missing Python Package

Fix

Network Connection Failed

Fix

GPU Driver Error

Fix

Process Terminated

Fix

Permission Denied

Frequently Asked Questions

What causes "GPU Memory Exhausted" in ComfyUI?

The model + activations exceed your GPU VRAM (e.g. an fp16 14B video model wanting ~28GB on a 24GB card).

How do I fix "GPU Memory Exhausted"?

Lower resolution first (1024×1024 → 832×1216 → 768×768) and set batch size to 1 — the cheapest fix. Use an fp8 / GGUF (Q5/Q4) quantized checkpoint — roughly halves weight VRAM at a small quality cost.

Which cloud GPU fixes this without changing my workflow?

Move to a larger GPU: a A6000 (48GB) runs from $0.64/hr and holds models a 24GB card cannot. The same workflow runs unchanged.