Most of the answers to "what GPU can run Stable Diffusion" online are a single number with no source. The project docs are more careful and less complete: some jobs have a published figure, some have none, and one of the best-known tables does not even label its units. This is what they say as of October 2026, and where you have to measure on your own card.
The goal here is a decision: can the card you have run the job you want, and which switch to flip when it cannot.
What you need before you start
- The name and memory of your GPU. On an Nvidia card,
nvidia-smi --query-gpu=name,memory.total --format=csvprints both. - Your front end: AUTOMATIC1111 WebUI, WebUI Forge or ComfyUI. The flags below differ between them.
- Which model family you plan to run. A Stable Diffusion 1.5 checkpoint and an 8B-parameter model have almost nothing in common as far as memory goes.
Step 1. Read the figures that exist
| Job | Figure in the docs | Note |
|---|---|---|
| Image, SD 1.x style, AUTOMATIC1111 | Peak 4.1 at batch 1, 6.2 at batch 2, out of memory at batch 4 (default attention). With xformers: 2.8 up to batch 4, 3.1 at batch 8 | AUTOMATIC1111 wiki, Optimizations; its table comes from an RTX 3060 and has no unit label |
| Image, lowest memory preset | Peak 0.9 at batch 1 with the lowvram preset | Same table, same caveat |
| ControlNet | "Low VRAM mode" recommended for 8GB GPUs or for bigger batches | ControlNet README |
| AnimateDiff, SDXL-Beta motion module | About 13GB VRAM for inference | AnimateDiff README; no figure for the SD 1.5 versions |
| ComfyUI, general | Can run on "as low as 4GB vram + 8GB ram" with asynchronous weight streaming | ComfyUI README |
| Stable Diffusion 3.5 Large | 8B parameters; a 4-bit NF4 setup with CPU offload is shown | Model card gives no VRAM number |
Source: AUTOMATIC1111 wiki, Optimizations for the first two rows; ComfyUI README for the ComfyUI row; the ControlNet, AnimateDiff and Stable Diffusion 3.5 Large pages are listed under the sources of this article.
Two cautions. The AUTOMATIC1111 wiki's numbers come from one RTX 3060 and CUDA 11.7, an old setup, so read them as a shape (how memory grows with batch size), not as a promise for a current card. And the ComfyUI line is a ceiling for how low it can go, not a recommendation: the README says "relatively quickly", with no timing.
Step 2. Cut memory in AUTOMATIC1111
The wiki lists the switches in order of how much they cost you.
--xformers: the wiki calls it a "Great improvement to memory consumption and speed", Nvidia GPUs only. Try this first.--medvram: splits the model into three parts so less sits in VRAM. The wiki says it lowers performance "only by a bit", except with live previews on.--lowvram: an "even more thorough optimization", and "Devastating for performance". Keep it for cards that fail everything else.
--opt-sdp-attention can be faster than xformers on some systems, but the wiki says it needs more VRAM, so it is the wrong pick on a tight card.
Step 3. Let ComfyUI and Forge manage it
ComfyUI's README lists smart VRAM and RAM management and model offloading among its features, so on a small card you often change nothing and it still runs, slowly. WebUI Forge exposes a "GPU Weight" setting that its README points to for Flux performance problems. Our notes on WebUI Forge cover where it came from.
Step 4. Plan for ControlNet and AnimateDiff
ControlNet adds layers on top of the base model, though its README says the required GPU memory "is not much larger than original SD". On 8GB, switch on the low VRAM mode the ControlNet README describes. For video, the only published number in the AnimateDiff README is the SDXL-Beta module's 13GB, so a card below 13GB would need tuning there. Our ControlNet plus AnimateDiff walkthrough shows the graph; keep resolution and frame count down on the first run, and raise them while watching memory.
Can it run on CPU?
Technically yes. The ComfyUI README says its Windows portable build works on Nvidia GPUs or on CPU only, and does not recommend it for regular users. It gives no timings, so there is no number to quote. Run one image and time it before committing to the setup.
What usually goes wrong
- Out-of-memory errors at the second or fourth batch item, not the first. Drop the batch size before changing anything else.
- Assuming a figure in one table carries to a different model. The wiki's table says nothing about SDXL or newer models.
- Turning on
--lowvramby default, then blaming the card for slow output.
If the docs give no figure for your case, the honest answer is to run one image, watch memory in nvidia-smi, and scale from there.
Who this suits
- A card with 4GB to 6GB: ComfyUI first, since its README states a 4GB floor, or AUTOMATIC1111 with
--xformers, then--medvram. Keep--lowvramas the last resort. - A card with 8GB: SD 1.x style images in AUTOMATIC1111 with
--xformers, and ControlNet with its low VRAM mode. - A card below 13GB and video work: expect tuning, because 13GB is the only AnimateDiff figure published, and it is for the SDXL-Beta module.
- Large models such as Stable Diffusion 3.5 Large: no VRAM number is published, so test first and use offload if the card runs out.