What runs on 16 GB of VRAM.

The 16 GB cards.

AMD’s 16 GB cards (RX 9070 XT, 9070, 7800 XT, 7600 XT) fit the same models: see the AMD guide.

What fits in 16 GB.

Fits runs from the card. Tight fits at the default size. Offloads keeps part of the model in system RAM: it works, slower.

• Krea 2 Turbo and Rawfp8 · 13.1 GBFits

  Or nvfp4 (7.7 GB) on RTX 50. The bf16 files (26.3 GB) offload.

• Flux.1 Devfp8 · about 12 GBFits

  16.7 s at 20 steps on a 5080, 6.7 s tuned.

• Flux.2 Klein 9Bfp8Fits

  Comfy measures 19.6 GB at bf16; fp8 fits.

• Z-Image, Qwen-Image 2.1, Chroma, MageFlow, Ideogram 44 to 14 GBFits

  Ideogram 4 in nvfp4 on RTX 50; its fp8 pair offloads a little.

• SD 3.5 Large, ERNIE-Image14.9 and 16.1 GBTight

  One image at a time at 1024 px.

• Qwen-ImageGGUF Q5_K_M · 14.9 GBTight

  fp8 (20.4 GB) offloads.

• HiDream I1fp8 · 17.1 GBOffloads

  A little over, plus four text encoders.

• Wan 2.2 5Bfp16 · 10 GBFits

  1280 × 704, 121 frames.

• HunyuanVideo 1.5fp8 · 8.3 GBFits

  720p.

• Wan 2.2 14Bfp8 · 2 × 14.3 GBOffloads

  Each half fits alone; the switch goes through RAM. 832 × 480 is easy, 1280 × 720 is slow.

• MiniMax H3w6a8 · 16 GBOffloads

  With its nvfp4 text encoder (15.7 GB). 64 GB of RAM.

• Flux.2 Devfp8 · 35.5 GBOffloads

  64 GB of RAM.

Set up ComfyUI.

1. 1

   Install or update ComfyUI

   ComfyUI Desktop or the portable NVIDIA build. RTX 50 cards need a PyTorch built for CUDA 12.8 or newer; the current portable ships CUDA 13.0.

2. 2

   Keep the default memory handling

   Since March 2026, ComfyUI manages graphics memory by itself on NVIDIA cards. It calls this Dynamic VRAM: model weights stream into the card as they’re needed, and the rest waits in system RAM without going to the pagefile. It’s on by default in current ComfyUI, on Windows and Linux (not WSL).

   That changes the old advice. --lowvram now does nothing while Dynamic VRAM is on, and --novram, --highvram and --gpu-only switch it off. --normalvram is gone, and an old shortcut that still passes it stops ComfyUI with unrecognized arguments.

   • --reserve-vram 1 keeps 1 GB free for Windows and the browser. Add it if the desktop stutters while ComfyUI renders.
   • --vram-headroom 1 asks Dynamic VRAM to keep extra room free, counting what other apps use.
   • --fast-disk prefers streaming weights from disk over RAM. ComfyUI says it can be faster with a fast NVMe SSD.
   • --disable-dynamic-vram brings back the old loader, for a custom node that breaks with the new one. ComfyUI says this flag will be removed soon.
3. 3

   Pagefile and RAM

   Models that don’t fit the card pass through system RAM, and Windows needs pagefile room while it loads them. Leave the pagefile on System managed size on your fastest SSD. If ComfyUI stops with The paging file is too small for this operation to complete, set a custom size of 32 GB or more under System › About › Advanced system settings › Performance › Advanced › Virtual memory.

   With 16 GB of VRAM, 32 GB of RAM covers the image models. Wan 2.2 14B, HiDream and Flux.2 Dev are happier with 64 GB.

Where 16 GB still runs out.

• Qwen-Image in fp8 is 20.4 GB. Use a GGUF Q4_K_M (13.1 GB) or Q5_K_M (14.9 GB), or let it offload.
• Krea 2 at full precision is 26.3 GB. The fp8 file is what fits.
• Wan 2.2 14B swaps two 14.3 GB halves through RAM, which is where 16 GB cards with 32 GB or less crash [4].
• Flux.2 Dev needs a 32 GB card or 64 GB of RAM.

Some 16 GB owners with 32 GB of RAM saw memory handling get worse after ComfyUI updates in 2026 [5]. If a workflow that used to work now stalls, update again first; if it persists, --disable-dynamic-vram brings back the old loader [6].

5070 Ti or 5080?

Same 16 GB, same formats. The 5080 has about 20% more cores and slightly more bandwidth, so it’s a bit faster at everything; it doesn’t fit anything the 5070 Ti can’t. For ComfyUI, the difference in price buys speed, not models.

When it goes wrong.

torch.OutOfMemoryError: CUDA out of memory

The model, the image size and the batch didn’t fit together. Generate one image at a time, go down a size, or use the fp8, int8 or GGUF file. When it fails at the very end, the VAE decode ran out: swap VAE Decode for VAE Decode (Tiled). For video, fewer frames help more than a smaller size.

The paging file is too small for this operation to complete. (os error 1455)

Windows ran out of virtual memory while loading a big file. Give the pagefile 32 GB or more, or add RAM.

It suddenly got much slower

The card is full and Windows is borrowing system RAM. Check Shared GPU memory in Task Manager. Close other GPU apps, lower the size, and set Prefer No Sysmem Fallback so it fails fast instead.

unrecognized arguments: --normalvram

An old launch flag. Remove it from the shortcut or the .bat file. Current ComfyUI doesn’t need memory flags on NVIDIA.

The first image takes minutes, the next ones don’t

The first run reads the model from disk. Later runs start warm. An SSD shortens the first one; with 16 GB of RAM it can stay slow, since the model gets pushed out between runs.

Questions.

Sources: [1] Flux Dev fp8 GPU benchmark thread, ComfyUI discussion #9002, [2] SDXL GPU benchmark thread, ComfyUI discussion #2970, [3] Krea 2 files, Comfy-Org on Hugging Face, [4] Wan 2.2 crashes after a GPU upgrade, ComfyUI issue #12451, [5] Memory regression on 16 GB cards, ComfyUI issue #12541, [6] Dynamic VRAM on by default, ComfyUI discussion #12699, [7] Dynamic VRAM in ComfyUI, Comfy blog, [8] Steam Hardware Survey, August 2026.