Diffusion Docs

Appearance

Out Of Memory ​

Out-of-memory errors happen when the workflow needs more GPU or system memory than is available.

Quick Take

Lower resolution, batch size, and heavy extras first. Those are usually the fastest memory wins.

Common Causes ​

  • image size too large
  • batch count too high
  • high-resolution fix or upscaling enabled
  • large model or multiple models loaded
  • memory-heavy extensions or custom nodes

Fixes ​

  • reduce width and height
  • set batch size to 1
  • disable upscaling or high-resolution passes
  • unload unused models
  • use memory-saving options supported by your tool
  • restart the UI if memory is fragmented or stuck

What Uses Memory ​

Memory pressure comes from several places:

  • model weights loaded into GPU memory
  • text encoders, VAEs, control models, and adapters
  • latent tensors for the current resolution and batch
  • attention operations inside the model
  • upscaling or high-resolution passes
  • previews, caches, extensions, or custom nodes

Large models, high resolutions, and multiple loaded components can combine quickly. A workflow that fits at 768x768 may fail at 1536x1536 because memory does not grow linearly in the way users expect.

Fastest Reductions ​

Try these before changing obscure launch flags:

  1. Set batch size to 1.
  2. Reduce width and height.
  3. Disable high-resolution fix, tiled upscale, or second-pass refinement.
  4. Remove extra control models and unused LoRAs.
  5. Close other GPU-heavy applications.
  6. Restart the UI to clear stuck allocations.

Advanced Options ​

Depending on the tool and hardware, memory-saving options may include attention slicing, tiled VAE decoding, CPU offload, lower precision, quantized models, or backend-specific optimizations. These can help, but they may slow generation or change compatibility.

Hardware Advice Is Tool-Specific

Exact memory options depend on the UI, backend, GPU, driver, and operating system. Follow tool-specific docs for flags and launch options.