Guides

The Launch-Flag Field Guide: Community Command-Line Hacks for Low VRAM

18 September 2026 · RenderBob team

Beyond block swap and GGUF, a second layer of VRAM control lives on ComfyUI's command line. Here is what the widely shared flags actually do, including where current defaults have moved.

Beyond the node-level techniques covered this week, block swap, GGUF, the Quantization Toolkit, a second layer of VRAM control lives at ComfyUI's command line. These flags apply globally rather than per-node. They are widely shared in troubleshooting threads, and some of them have become defaults, so the useful move is knowing what each one still does rather than pasting an old forum line unchanged.

--fp8_e4m3fn-unet and --fp8_e4m3fn-text-enc force FP8 encoding on the diffusion model and text encoder at launch, independent of precision settings inside the workflow. Community testing confirms this works with an otherwise unmodified, full-precision workflow loaded. Given the FP8 silent-upcast problem earlier this week, these flags are the more forceful, launch-level alternative when a node-level setting is not producing the VRAM reduction it should.

--lowvram is the longer-standing flag. On current ComfyUI it does nothing when dynamic VRAM is already enabled; otherwise it keeps text encoders on CPU. Reach for it on a constrained card before the more specific flags, and check whether dynamic VRAM is already doing the job. --async-offload overlaps weight movement between CPU and GPU with compute rather than blocking it. On NVIDIA it is already enabled by default, with an optional stream count; --disable-async-offload turns it off. --preview-method none disables live step previews. Previews cost VRAM and bandwidth during iteration-heavy sessions. Current ComfyUI defaults this to none, so the useful case is turning previews off again after someone enabled latent2rgb or taesd.

A Windows-managed page file is not a ComfyUI flag. It is standard community guidance alongside these: keep the system page file system-managed rather than disabled or fixed too small, so the OS has a release valve when RAM, not just VRAM, gets tight. Several low-VRAM workflow guides list this early, because RAM pressure causes a surprising share of out-of-memory failures on constrained machines.

Pin the flags this studio's low-VRAM nodes launch with by default, and the first three things to add when a new machine or a new model starts throwing OOM errors. That turns a six-month-old forum post into a five-minute, repeatable fix.