Skip to content

How-ToRAM

How to Run ComfyUI on a Low-VRAM Graphics Card: The Settings That Actually Matter

ComfyUI now manages graphics memory itself with dynamic VRAM and async offloading, and its documentation says --lowvram does nothing while dynamic VRAM is on. On a 6GB or 8GB card, start with no flags, then use --reserve-vram or --vram-headroom if the rest of the PC struggles and --novram as a last resort. Smaller, FP8 or quantised models and lower resolutions save the most memory. Forge PC has no

By , Editorial team

4 min read

How to Run ComfyUI on a Low-VRAM Graphics Card: The Settings That Actually Matter

If you search for how to run ComfyUI on a card with 6GB or 8GB of VRAM, most guides tell you to add --lowvram. That advice is now often out of date. Current versions of ComfyUI manage graphics memory automatically, and in many setups --lowvram does nothing at all.

Forge PC has not tested these settings for this guide. The options below are taken from ComfyUI's own README and its command-line argument definitions, checked on 15 November 2026. Option names can change between releases, so check them against your version.

What ComfyUI does by default

ComfyUI's README says it can run "even the biggest open source models on as low as 4GB vram + 8GB ram relatively quickly" using asynchronous weight streaming, and lists smart VRAM and RAM management, model offloading and support for quantised models as standard features. That is ComfyUI's claim, not our measurement, but it explains why you rarely need to tune anything to get started.

Two defaults do most of the work:

  • Dynamic VRAM. ComfyUI uses dynamic VRAM management unless you turn it off or choose one of the fixed modes (--highvram, --gpu-only, --novram or --cpu).

  • Async offload. Asynchronous weight offloading is enabled by default on NVIDIA cards, so model weights are streamed in from system RAM rather than all loaded at once.

The practical upshot: on a low-VRAM card, system RAM matters too, because offloaded model weights have to sit somewhere. The more system memory you have, the more room ComfyUI has to offload. Our PC for Stable Diffusion and Flux guide covers the hardware side.

Step 1: Start with no special flags

Install ComfyUI the normal way (our ComfyUI install guide covers Windows), load a workflow and run it before changing anything. The README recommends the desktop app for regular users; the portable Windows build is aimed at NVIDIA GPUs or CPU-only use, and ComfyUI says to update your NVIDIA driver if the portable build doesn't start.

While it runs, watch dedicated GPU memory and system RAM using the method in our guide to checking VRAM usage. If generation completes and the PC stays responsive, leave the settings alone.

Step 2: Know which launch options do what

If you start ComfyUI from the command line (for example python main.py), you can add options to the end of the command. These are the memory-related ones, with ComfyUI's own descriptions:

Option

What ComfyUI says it does

When to try it

--lowvram

Does nothing if dynamic VRAM is enabled; otherwise runs the text encoders on the CPU

Only if you have disabled dynamic VRAM

--novram

"When lowvram isn't enough"

Last resort before CPU-only

--cpu

Uses the CPU for everything (slow)

Testing, or no usable GPU

--reserve-vram 1

Reserves the given number of GB of VRAM for Windows and other apps

Desktop stutters or other apps crash while generating

--vram-headroom 1

Keeps extra VRAM free under dynamic VRAM, counting other apps

Same as above, with dynamic VRAM on

--disable-smart-memory

Forces aggressive offloading to system RAM instead of keeping models in VRAM

Switching between several models

--cache-none

Cuts RAM and VRAM use, but re-runs every node each time

Very tight on system RAM

--disable-dynamic-vram

Goes back to estimate-based model loading

Troubleshooting only

The 1 values are examples; set the number of gigabytes that suits your system. Note that --novram and --cpu also switch dynamic VRAM off.

Step 3: Use smaller model files

Launch options can only move work around. The bigger saving usually comes from the model itself:

  • Choose a lighter model for the job. ComfyUI supports everything from Stable Diffusion 1.5 and SDXL to much larger models such as Flux.2. Older, smaller models are far easier on a 6GB or 8GB card.

  • Use quantised or FP8 versions. ComfyUI supports quantised models, and it has options to store the diffusion model (--fp8_e4m3fn-unet) and text encoder (--fp8_e4m3fn-text-enc) weights in FP8. Many model pages also offer pre-quantised files, which are smaller to download as well.

  • Lower the resolution and batch size first when you hit out-of-memory errors. These are workflow settings, not launch options.

Step 4: Free up VRAM outside ComfyUI

Other apps use VRAM too; ComfyUI's own --vram-headroom option exists because it has to count VRAM used by other software. Close what you don't need while generating. If you are still short, our VRAM guide explains what each capacity class is realistically good for.

Our verdict

On a low-VRAM card, start ComfyUI with no special flags and let dynamic VRAM and async offloading do their job. Only add options when you have a specific problem: --reserve-vram or --vram-headroom if the rest of the PC struggles, and --novram as a last resort. Smaller, quantised or FP8 models and lower resolutions will save more memory than any flag, and plenty of system RAM gives offloading more room. Forge PC has not tested these settings; descriptions are from ComfyUI's documentation.

The bottom line

On a low-VRAM card, start ComfyUI with no special flags and let dynamic VRAM and async offloading do their job. Only add options when you have a specific problem: `--reserve-vram` or `--vram-headroom` if the rest of the PC struggles, and `--novram` as a last resort. Smaller, quantised or FP8 models and lower resolutions will save more memory than any flag, and plenty of system RAM gives offloading more room. Forge PC has not tested these settings; descriptions are from ComfyUI's documentation.

Sources

  1. 01ComfyUI, README
  2. 02ComfyUI, command-line arguments (cli_args.py)

About the author

Forge PC Staff

Editorial team

The shared byline for articles researched, written and edited by the Forge PC editorial team. How we source, check and correct our work is set out in our editorial policy.

More from Forge PC Staff

Read next

How to Enable AMD Fluid Motion Frames (AFMF 2.1) on a Radeon Graphics Card

How-To

How to Enable AMD Fluid Motion Frames (AFMF 2.1) on a Radeon Graphics Card

To enable AMD Fluid Motion Frames, install Adrenalin 25.3.1 or newer, then select the HYPR-RX profile (RX 7000 and newer) or turn on AMD Fluid Motion Frames under Gaming > Graphics, globally or per game. AFMF 2.1 works on RX 6000 cards and newer in DirectX 11, DirectX 12, Vulkan and OpenGL games, with VSync off; RX 6000 cards need exclusive full screen.

Forge PC Staff2 min

Two vs Four RAM Sticks for DDR5: Why 2x32GB Usually Beats 4x16GB

Versus

Two vs Four RAM Sticks for DDR5: Why 2x32GB Usually Beats 4x16GB

Two DDR5 sticks are better for most builds. AMD lists DDR5-5600 as the AM5 maximum with two modules but DDR5-3600 with four, and XMP or EXPO speeds above that are overclocking and not guaranteed. Buy 2x16GB or 2x32GB for 32GB or 64GB, and use four sticks, ideally a matched kit, only if you need 128GB or more.

Forge PC Staff3 min

AMD's FSR 4.1.2 Unreal Engine 5 Plugin Brings ML Upscaling to RX 6000 Cards

News

AMD's FSR 4.1.2 Unreal Engine 5 Plugin Brings ML Upscaling to RX 6000 Cards

AMD's FSR Unreal Engine 5 plugin now uses FSR Upscaling 4.1.2, bringing ML-based upscaling to Radeon RX 6000 cards and select Ryzen integrated graphics as well as RX 7000 and RX 9000, in UE 5.3 to 5.8 games that adopt it. It's DirectX 12 only for now and adds Frame Generation 4.0.2 with driver-assisted frame pacing on RX 9000 cards.

Forge PC Staff2 min

NVIDIA's DGX Spark HPC Developer Kit Links Four Sparks Into a 512GB Desk-Side Cluster

News

NVIDIA's DGX Spark HPC Developer Kit Links Four Sparks Into a 512GB Desk-Side Cluster

NVIDIA's DGX Spark HPC Developer Kit, announced at SC26 on 16 November 2026, links four DGX Spark systems in a ring over ConnectX-7 with no network switch, for 512GB of combined memory and 20TB of NVMe storage. NVIDIA says it can spread models of up to 700B parameters across the nodes with NVFP4. No UK kit price yet; Scan sells two-Spark bundles at about £12,845.

Forge PC Staff3 min