ComfyUI VRAM requirements

Peak VRAM for all 11 models

At 1024×1024, batch 1, with the transformer, text encoder and VAE all resident in VRAM. Precision is the lever that matters most: it scales the weights, which dominate everything else at these model sizes.

“Lowest possible” is NF4 / INT4 with whichever offload mode is cheapest for that model — and it is not always the most aggressive one. Models with a very large text encoder do better by offloading the encoder alone than by cycling every component through VRAM, because sequential mode still has to hold the largest single component. It is how a 32B model reaches 22.5GB, and it is slow either way, because each step waits on the PCIe bus rather than on the GPU. For a specific resolution, batch size or offload mode, the image generation VRAM calculator breaks the same numbers down component by component.

The three ways to fit a model

When a model does not fit, ComfyUI does not refuse — it changes what it keeps resident. These are the modes, in order of how much speed they cost you.

• Everything in VRAM

  Transformer, text encoders and VAE all resident. Fastest, and the most memory-hungry.

• Text encoder on CPU

  Encode the prompt on CPU, then keep only the transformer and VAE in VRAM. Costs a couple of seconds per prompt and saves a lot on Flux and SD 3.5.

• Sequential CPU offload

  Stream components in and out of VRAM one at a time. Dramatically lower VRAM, dramatically slower — often the only way to run a 12B model on an 8GB card.

Precision is the other lever, and usually the better one to reach for first: BF16 / FP16 reference quality. what the repos ship by default, FP8 roughly half the weight memory. small quality cost, widely used for flux and sd 3.5, and NF4 / INT4 aggressive 4-bit. visible quality loss, but it puts big models on small cards.

Workflows, nodes and the Manager

ComfyUI or Automatic1111?

Running FLUX in either one raises a second question — which FLUX. FLUX.1 [dev] vs FLUX.1 [schnell] covers the licence difference as well as the step count, and the licence is the part that catches people out.

ComfyUI alternatives

Every front end below runs the same open checkpoints, and the VRAM figures above apply to all of them — the model sets the memory, not the interface. What differs is how you work, and whether the project is still keeping up with new architectures.

• SwarmUI

  Active · MIT

  The closest thing to ComfyUI without the graph. SwarmUI can install ComfyUI and use it as its generation backend, so it inherits ComfyUI's model support, and it puts a conventional generate screen in front — with the full node editor still one tab away when you need it.

• InvokeAI

  Active · v6.14.1, Sep 2026 · Apache 2.0

  Built around a canvas for inpainting, outpainting and layered editing rather than around a pipeline. The pick if your work is refining images by hand more than batch-generating them.

• AUTOMATIC1111

  Last tagged release v1.10.1, Feb 2025 · AGPL-3.0

  The form-based web UI that most older tutorials assume. Enormous extension ecosystem and still usable for SD 1.5 and SDXL, but development has slowed, and new architectures reach it late or not at all.

• Forge

  Last commit Jul 2025 · AGPL-3.0

  lllyasviel's fork of AUTOMATIC1111, built to manage memory better and run faster in the same familiar interface. It picked up Flux, which A1111 itself does not support, though it has not been updated since mid-2025.

• Fooocus

  Bug fixes only · GPL-3.0

  The simplest of the group — a prompt box with sensible defaults. Its own README describes it as in limited long-term support with bug fixes only and no plans for newer architectures, and points Flux users to Forge or ComfyUI/SwarmUI.

Release and commit dates read from each project’s GitHub repository on 15 September 2026.

Where to go from here

Common questions

How much VRAM does ComfyUI need?

ComfyUI's own footprint is small — the memory goes almost entirely to the model you load. At 1024x1024 and BF16 with everything resident, SDXL peaks near 9GB while FLUX.1 [dev] needs over 36GB. Drop to FP8 and FLUX comes down to roughly half that. The table on this page gives the figure for every model at each precision, and ComfyUI is unusually good at running models that do not fit: it will offload the text encoder, or move components through VRAM one at a time, at a cost in speed.

Is ComfyUI better than Automatic1111?

They serve different working styles rather than competing on quality — the same checkpoint produces the same image in both. ComfyUI exposes the generation pipeline as a node graph you wire up yourself, which is why it gets support for new architectures first and why it handles low-VRAM cards better. A1111 presents a form with fields, which is faster to start with and harder to extend. If you are chasing a specific multi-stage pipeline, ComfyUI; if you want to type a prompt and press generate, A1111.

What is the best alternative to ComfyUI?

If you want ComfyUI's model support without wiring node graphs, SwarmUI — it installs and drives ComfyUI as its backend and puts a simpler generate screen in front of it. InvokeAI is the actively developed option built around a canvas. AUTOMATIC1111 and Forge still work, but A1111's last tagged release was February 2025 and Forge's last commit July 2025, and Fooocus is in bug-fix-only support with no plans for newer architectures such as Flux.

What is a ComfyUI workflow?

A workflow is the node graph itself — the wiring that takes a checkpoint, a prompt, a sampler and a VAE and produces an image. ComfyUI saves it as JSON, and it also embeds the graph in the PNG metadata of every image it generates. That is the part people underestimate: dragging someone's ComfyUI output onto the canvas rebuilds the exact graph that made it, which is why workflows circulate as images rather than as files.

Can ComfyUI run on 8GB of VRAM?

Yes, and the limits are tighter than the folklore suggests. At 1024x1024 an 8GB card holds SD 1.5 at full precision with room to spare, and SDXL once you drop to FP8 — at BF16 SDXL already wants more than 8GB. SD 3.5 Medium fits only by offloading its text encoder, which costs less than you would expect because that encoder is the largest thing in the model. The 12B FLUX models bottom out near 10GB even at 4-bit with offload, so at native resolution they are genuinely out of reach on 8GB; generating at 512x512 brings them within it, because activations scale with pixel count.

Do I need ComfyUI Manager?

Not to generate images, but almost everyone installs it. Custom nodes are distributed as Git repositories that drop into custom_nodes/, and Manager handles finding, installing and updating them, plus flagging the missing nodes when you load someone else's workflow. Treat custom nodes as what they are — third-party code that runs with ComfyUI's permissions — and install from sources you have reason to trust.

Picking a card for image work, what each VRAM tier buys, and how memory is sized.

• How to Run Stable Diffusion Locally: Hardware, VRAM and SetupCurrent models from SD 1.5 to Qwen-Image, and the VRAM each needs.Read the guide
• ComfyUI Templates: 660 Built In, and Half Need the CloudWhat ships, what is cloud-only, and what each one downloads.Read the guide
• Local AI Video Generation in ComfyUIWan 2.2, HunyuanVideo and LTX-2.5, with vendor memory figures.Read the guide