ComfyUI Ops: Node Graphs, SDXL/Flux VRAM and OOM Fixes

Start here

ComfyUI exposes a diffusion pipeline as an explicit node graph rather than a form with a prompt box. That trade buys direct control over every stage and costs you a set of operational problems that do not exist in simpler interfaces: memory ceilings that move with resolution and batch size, a custom node ecosystem that installs into a shared Python environment, and a graph whose caching behaviour decides how fast you can iterate.

This site covers those problems, sourced from project documentation, published model weight sizes and the runtime's own command line reference rather than from anecdote.

• ComfyUI GPU requirements: how much VRAM you need — published weight sizes for SD 1.5, SDXL and Flux mapped onto VRAM tiers, plus the system RAM and disk that go with them.
• ComfyUI out of memory errors: causes and fixes — reading the allocator message, telling loader failures from decode failures, and the flags that lower peak usage.
• ComfyUI vs Automatic1111 vs Forge — release cadence, memory architecture and extension risk across the three main local interfaces.
• How ComfyUI graphs execute and where VRAM goes — pull-based execution, per-node caching, and the three places memory pressure comes from.

For a figure rather than a tier, the ComfyUI VRAM and execution sizer estimates the footprint of a specific checkpoint class, adapter count, resolution and batch size.