Qwen-Image-2.1 VRAM Calculator

Common setups at 1024×1024

Low-VRAM setups for 8, 12 and 16 GB

Each row is a mix of the files above at 1024×1024. The three phase columns are the weights on the GPU while the prompt is encoded, while the image is denoised and while the VAE decodes it: exact file sizes. The peak adds the working memory measured in real runs, 1.6–2.4 GB with the text encoder on the CPU.

On Windows, “runs on 8 GB” can mean 18 GB

NVIDIA’s Windows driver moves what does not fit into shared system memory instead of failing (System Memory Fallback, since driver 536.40), so a job finishes but runs from system RAM. One RTX 3070 8GB owner saw the INT8 set use 7.3 GB of dedicated VRAM plus 11.2 GB of shared memory. On the same card, the Q4_K_M DiT on the GPU with the Q4_K_M encoder on the CPU peaked at 6.9 GB, and Q3_K_M files for both, together 6.8 GB, ran out of memory while decoding. If Task Manager shows “Shared GPU memory” climbing during a run, it does not fit; setting “CUDA - Sysmem Fallback Policy” to “Prefer No Sysmem Fallback” in the NVIDIA Control Panel makes that an out-of-memory error instead.

On a Mac (unified memory)

A public report from an M5 Max 48 GB shows Unsloth Desktop on macOS using about 34 GB after loading and about 42 GB while generating with the Q4_K_M DiT (image size not stated), far more than the CUDA figures on this page. The load figure is close to the files with the DiT held at BF16 (32.4 GB with a BF16 encoder and VAE, against 22.4 GB with the Q4_K_M DiT kept quantized, in decimal GB), so that path likely dequantizes it; this has not been confirmed. The ranges here come from CUDA runs (ComfyUI and diffusers) and may underestimate Mac paths such as MPS or MLX. macOS also lets the GPU use only part of the RAM by default, logged at about two thirds to three quarters, so leave headroom on a 48 GB Mac; in the same thread, the user’s 8-bit MLX build in MLX Serve peaked at about 20 GB. stable-diffusion.cpp runs GGUF files on Metal, but we have no Mac measurement of it.

Settings that do this

• ComfyUI: --disable-dynamic-vram --lowvram runs the text encoders on the CPU (with dynamic VRAM on, --lowvram does nothing); the “VAE Decode (Tiled)” node decodes in tiles; --reserve-vram 1 keeps 1 GB free for the desktop; --novram is the last resort when --lowvram is not enough.
• stable-diffusion.cpp: --backend te=cpu runs the text encoder on the CPU (it replaces the deprecated --clip-on-cpu); --vae-tiling decodes in tiles; --diffusion-fa uses flash attention in the DiT; --offload-to-cpu keeps the weights in system RAM and copies them to the GPU when needed; --max-vram cuda0=6 caps what it plans to use.

Flags checked on 2026-09-29 in ComfyUI’s cli_args.py and nodes.py, and stable-diffusion.cpp’s performance, backend and Qwen Image 2.1 docs.

Peak VRAM people measured

File sizes read from Hugging Face on 2026-09-28: Comfy-Org/Qwen-Image-2.1, unsloth/Qwen-Image-2.1-GGUF, unsloth/Qwen-Image-2.1-FP8, gguf-org/qwen-image-2.1-gguf, unsloth/Qwen3-VL-8B-Instruct-GGUF, Qwen/Qwen-Image-2.1. For language models, see the LLM VRAM calculator. How the site’s estimates compare with real runs: predicted vs measured.