Short answer
Count the unified memory, then take about 70%: that’s what the GPU can use comfortably. 16 GB runs SDXL, 18 to 24 GB adds Flux.2 Klein and Z-Image, 48 GB runs Krea 2 at full precision, and 128 GB runs Flux.2 Dev.
It works, and it’s slower than an NVIDIA card. The same SDXL image took about 40 seconds on an M1 Max and 16 on an RTX 4060 Laptop.
How much memory the GPU gets.
A Mac shares one memory between CPU and GPU, and macOS lets the GPU use part of it. Metal calls that part the recommended working set: on a 24 GB M4 Pro it’s 19.07 GB, about 74%, and reports elsewhere put it at 75 to 78% [1]. ComfyUI itself reports the whole memory, so it can’t tell you. Plan with about 70%, which leaves macOS and your other apps some room.
| Memory | About 70% for the GPU | Runs well | HEISS UI marks |
|---|---|---|---|
| 8 GB | 5.6 GB | SD 1.5, SDXL Lightning at a squeeze | Nothing |
| 16 GB | 11.2 GB | SDXL, Z-Image as a GGUF, Flux.2 Klein 4B tightly | SDXL |
| 18 to 24 GB | 12.6 to 16.8 GB | Flux.2 Klein 4B, Z-Image, Flux.1 Dev as a GGUF Q4 to Q6 | Flux.2 Klein 4B, SDXL |
| 32 to 36 GB | 22 to 25 GB | Qwen-Image 2.1, Flux.1 Dev GGUF Q8, Wan 2.2 5B, HunyuanVideo 1.5 | Flux.2 Klein 4B, SDXL |
| 48 to 64 GB | 34 to 45 GB | Krea 2 at full precision, Qwen-Image, Wan 2.2 14B in GGUF halves | Krea 2 Turbo, Flux.2 Klein 4B, SDXL |
| 96 to 128 GB | 67 to 90 GB | Flux.2 Dev, MiniMax H3, Wan 2.2 14B at full precision | Krea 2 Turbo; Flux.2 Dev at 128 GB |
Raising the limit.
On macOS 14 and later, iogpu.wired_limit_mb sets how much the GPU may use. It resets at restart. Leave several GB for macOS: the example gives a 32 GB Mac 26 GB for the GPU.
sudo sysctl iogpu.wired_limit_mb=26624HEISS UI reads this value when ComfyUI runs on the same Mac, and counts it instead of the 70%.
Skip fp8 files on a Mac.
Apple’s GPU can’t compute in fp8, and ComfyUI loads fp8 weights as fp16 or bf16 on a Mac. An fp8 download saves disk space, then takes as much memory as the full-precision file [1]. Some fp8 files don’t load at all and stop with Trying to convert Float8_e4m3fn to the MPS backend [5].
- To make a model smaller on a Mac, use a GGUF (Q4_K_M to Q8_0) through the ComfyUI-GGUF nodes.
- Otherwise take the bf16 or fp16 file.
- The int8 and nvfp4 files are tuned for NVIDIA cards; stick to GGUF and bf16 on a Mac.
Set up ComfyUI on a Mac.
Install ComfyUI Desktop
It needs Apple Silicon (M1 or later) and macOS 13 or newer [8]. For a manual install, ComfyUI’s README asks for the latest PyTorch nightly on Apple Silicon.
Close what uses memory
Browsers with many tabs, video apps and other AI tools share the same memory as the GPU. On 16 and 24 GB Macs, quitting them is the cheapest speed-up there is.
Download GGUF or bf16 files
Not fp8, for the reasons above. For Flux.1 Dev on 24 GB, a GGUF Q5_K_S (8.3 GB) or Q6_K (9.9 GB) is a good balance.
Start small
Generate at 1024 px, one image at a time. For video, start at 320 × 320 to check it runs at all, then go up [7].
How long it takes.
| Mac | Model and settings | Time |
|---|---|---|
| Mac mini M4 Pro, 24 GB | SDXL, 1024 px, 25 steps | 20 to 40 s[3] |
| M1 Max | SDXL, 1024 px, 20 stepsThe same image took 15.8 s on an RTX 4060 Laptop. | about 40 s[2] |
| M1 Max, 64 GB | Flux.2 Klein 4B, 1024 px, 4 stepsIn mflux (MLX), not ComfyUI. | 31.7 s[4] |
There are no reliable ComfyUI numbers for Krea 2, Qwen-Image or video on current Macs yet.
When it goes wrong.
MPS backend out of memory (MPS allocated: … max allowed: …)- The model and the image didn’t fit the GPU’s share. A 16 GB Mac mini hit this with SD 3.5 Large [6]. Use a smaller file or size, and quit other apps. The message suggests
PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0; that removes the limit, and the Mac can freeze instead. Trying to convert Float8_e4m3fn to the MPS backend- An fp8 file. Download the bf16 or a GGUF version instead.
- A Flux image takes many minutes
- Usually an fp8 file expanded to full size, or the Mac swapping to disk. Check Activity Monitor › Memory for swap. Use a GGUF that fits your memory.
- Video comes out black, or stalls
- Memory pressure. Lower the size and the frame count.
Where other apps are faster.
ComfyUI runs everything on this page, and it isn’t the fastest option on a Mac. Draw Things uses Metal-specific attention and was about 20% faster than ComfyUI on the same Mac mini [3]. mflux runs Flux models through Apple’s MLX. If you only need one model they support, they’re worth a look.
A Mac and a PC.
If there’s an NVIDIA PC in the house, run ComfyUI there and work from the Mac. Start ComfyUI on the PC with --listen, and point the Mac’s front end at the PC’s address.
Questions.
How much memory do I need on a Mac for AI images?
16 GB runs SDXL. 24 GB is comfortable for Flux.2 Klein, Z-Image and Flux.1 Dev as a GGUF. 48 GB runs Krea 2 at full precision, and 128 GB runs Flux.2 Dev.
Why is my Mac so much slower than a PC?
Apple’s GPU does less work per second than a desktop NVIDIA card, and ComfyUI is tuned for NVIDIA first. An SDXL image took about 40 seconds on an M1 Max and 16 on an RTX 4060 Laptop.
Should I use fp8 files on a Mac?
No. They load at full precision, so they save disk space but not memory, and some don’t load at all. Use a GGUF or the bf16 file.
What is iogpu.wired_limit_mb?
A macOS 14+ setting for how much memory the GPU may use. Raising it with sudo sysctl lets bigger models run; it resets at restart. Leave several GB for macOS.
M4 Pro with 24 GB or M4 Max with 36 GB?
For image models, the extra memory matters more than the chip: 36 GB runs Qwen-Image 2.1 and larger GGUFs of Flux.1 Dev comfortably. The Max also has a bigger GPU.
Can a Mac make video?
Yes, slowly. Wan 2.2 5B and HunyuanVideo 1.5 run on 32 GB and up, Wan 2.2 14B as GGUF halves on 64 GB. Start at a small size to check it runs.
Sources: [1] HEISS UI hardware notes, [2] SDXL on an RTX 4060 Laptop, lilting channel, [3] Local image generation on a Mac mini M4 Pro, heyuan110, [4] Flux.2 Klein 4B on an M1 Max, lilting channel, [5] fp8 files on Apple Silicon, ComfyUI discussion #13273, [6] MPS backend out of memory on a 16 GB Mac mini, ComfyUI issue #7171, [7] Wan 2.2 on Apple Silicon, Papaya Bytes, [8] ComfyUI Desktop for macOS, Comfy docs.
