What runs at each size.
Memory decides what runs; speed decides how long it takes. From the card means the model fits in graphics memory. With system RAM means ComfyUI keeps part of it in RAM and streams it in, which works, slower, with 32 GB or more.
| Memory | From the card | With system RAM | Guide |
|---|---|---|---|
| 6 GB | SD 1.5, SDXL Lightning, Z-Image as a GGUF | Flux.2 Klein 4B, Wan 2.2 5B | Laptops |
| 8 GB | SDXL, Flux.2 Klein 4B, Z-Image, MageFlow, Wan 2.2 5B | Krea 2, Flux.1 Dev, Qwen-Image 2.1, Wan 2.2 14B | 8 GB, 3060 Ti |
| 12 GB | Adds Flux.2 Klein at full precision, Qwen-Image 2.1, Chroma, Krea 2 just about | Wan 2.2 14B, Qwen-Image, HiDream, MiniMax H3 | 12 GB, 3060 12 GB |
| 16 GB | Adds Krea 2, Flux.1 Dev, SD 3.5 Large, 720p video | Qwen-Image fp8, HiDream, Wan 2.2 14B, Flux.2 Dev | 16 GB, 5060 Ti |
| 24 GB | Adds Qwen-Image, HiDream, MiniMax H3 pruned, Wan 2.2 14B | Flux.2 Dev, Krea 2 at full precision | 3090 and 4090 |
| 32 GB | Nearly everything at full precision | Flux.2 Dev in fp8, a little | 5090 |
Steam, August 2026: 25.7% of users have 8 GB, 13.0% have 12 GB, 26.9% have 16 GB and 5.4% have 24 GB.
System memory matters too.
Once a model doesn’t fit the card, RAM decides whether it runs well. 41% of Steam users have 16 GB [1], and that’s the most common reason a big model takes minutes to load: one RTX 3060 owner with 16 GB waited 10 to 20 minutes for Flux.1 Dev [6].
| Graphics memory | 16 GB RAM | 32 GB RAM | 64 GB RAM |
|---|---|---|---|
| 8 GB | Fits-only | Most models | Nearly all |
| 12 GB | Fits-only | Most models | Nearly all |
| 16 GB | Images | Images, most video | Everything but Flux.2 Dev fp8 |
| 24 GB | Most models | Nearly all | Everything |
“Fits-only” means the models that fit the card. Everything else loads slowly and leans on the pagefile.
Three things besides memory.
The card’s generation.
RTX 40 and 50 cards compute fp8 directly; RTX 50 cards also compute nvfp4, a 4-bit format Comfy-Org now ships for Krea 2, Z-Image, Ideogram 4 and Qwen-Image [7]. RTX 30 cards, Radeons and Macs run fp8 files too, but only to save memory: on a Mac they don’t even do that [5].
ComfyUI’s memory handling.
Since March 2026 ComfyUI streams model weights into NVIDIA cards as it needs them, which it calls Dynamic VRAM [2]. The old --lowvram flag does nothing while it’s on [3]. On AMD it only turns on with recent ROCm builds.
Windows’ shared GPU memory.
When the card is full, the NVIDIA driver on Windows can borrow system RAM instead of failing, and a render crawls [4]. Setting Prefer No Sysmem Fallback for ComfyUI’s python.exe in the NVIDIA Control Panel turns that into a clear error. Every NVIDIA guide here walks through it.
Questions.
How do I find out how much VRAM I have?
On Windows, Task Manager › Performance › GPU shows Dedicated GPU memory. With an NVIDIA card, nvidia-smi in a terminal shows it too. On a Mac, Apple menu › About This Mac shows the memory; the GPU can use about 70% of it.
Does system RAM count as VRAM?
No, but ComfyUI can use it to hold the parts of a model that don’t fit the card, and stream them in. That works, and it’s slower. On a Mac, the memory is shared, and the GPU gets most of it.
Is 8 GB of VRAM enough in 2026?
For SDXL, Flux.2 Klein 4B, Z-Image and Wan 2.2 5B, yes. Bigger models run too, with 32 GB of system RAM, more slowly.
Do I still need --lowvram?
Not on an NVIDIA card with a current ComfyUI: it manages memory by itself, and --lowvram does nothing while that’s on. On AMD, AMD’s own guide still uses it on 12 GB cards.
Is a Mac or an NVIDIA card faster?
An NVIDIA card. The same SDXL image took about 40 seconds on an M1 Max and 16 on an RTX 4060 Laptop. A Mac with a lot of memory can load bigger models than most graphics cards, slowly.
Sources: [1] Steam Hardware Survey, August 2026, [2] Dynamic VRAM in ComfyUI, Comfy blog, [3] ComfyUI launch flags, cli_args.py, [4] System memory fallback, NVIDIA, [5] HEISS UI hardware notes, [6] Flux on an RTX 3060 with 16 GB of RAM, ComfyUI issue #12334, [7] Krea 2 files, Comfy-Org on Hugging Face, [8] SDXL GPU benchmark thread, ComfyUI discussion #2970.
