Check my GPU →

Why the first image in ComfyUI takes longer: 13 files timed

First run vs the runs after it, 13 files on an RTX 5060 Ti 16 GB: most native files added 3 to 7 seconds, every GGUF Q8_0 file 10 or more.

TimedBy Tomasz · aboutUpdated 2026-11-17

Everyone notices it: the first image after starting ComfyUI, or after switching models, takes longer than the next ones. I had the numbers anyway, because every test I run starts with one warm-up run before the timed ones. So here is how much longer the first run was, file by file.

My setup: RTX 5060 Ti 16 GB, 32 GB of RAM, ComfyUI portable, 1024×1024 images (the two Kontext rows are edits of a 1024×1024 picture). "First run" means the run right after the models were unloaded, so it includes reading the files into memory and encoding the prompt. "Next runs" is the average of the timed runs that came after it, with a new seed each time.

ModelFileFirst runNext runsExtra
FLUX.1 schnellGGUF Q8_022.8 s9.4 s+13.4 s
FLUX.1 devFP839.1 s36.4 s+2.7 s
FLUX.1 devGGUF Q8_050.7 s40.6 s+10.1 s
FLUX.1 devGGUF Q4_K_S52.1 s47.3 s+4.8 s
FLUX.1 KontextFP860.0 s53.4 s+6.6 s
FLUX.1 KontextGGUF Q8_094.1 s82.1 s+12.0 s
Qwen-Image 2.1INT824.6 s17.2 s+7.4 s
Qwen-Image 2.116-bit46.2 s39.7 s+6.5 s
Qwen-Image 2.1GGUF Q8_053.9 s43.6 s+10.3 s
Qwen-ImageGGUF Q4_K_M138.7 s123.3 s+15.4 s
Z-Image Turbo16-bit15.4 s11.0 s+4.4 s
FLUX.2 klein 4BFP86.7 s2.9 s+3.8 s
Ming-ImageINT817.6 s7.4 s+10.2 s

What I see in it

The extra time is seconds, not minutes. For the FP8, INT8 and 16-bit files it was mostly 3 to 7 seconds. Once a model is loaded, ComfyUI keeps it in memory, and only the first image pays.

The GGUF Q8_0 files paid more. With FLUX.1 dev, the FP8 file added 2.7 seconds and the Q8_0 GGUF 10.1. FLUX.1 Kontext: 6.6 against 12.0. Qwen-Image 2.1: 7.4 for INT8, 6.5 for 16-bit, 10.3 for Q8_0. Every Q8_0 file I tested added 10 seconds or more. I do not know the exact reason. My guess is the extra work the GGUF loader does when it reads the file, but I have only measured it, not traced it.

Ming-Image is the exception. Its INT8 file is fast, yet the first run took 10 seconds longer. Its text encoder is a big language model, and loading it is most of the wait.

When it matters

If you make a batch of images with one model, not at all: you pay once. It matters when a workflow switches between models, or when you jump between models all evening. Then the native files (FP8, INT8, 16-bit) save you a few seconds every time, on top of being faster per image on an RTX 40 or 50 card. FP8 vs GGUF per image →

If the first run takes minutes instead of seconds, that is a different problem: usually RAM is full and Windows is using the page file. How much RAM you need →

Every run behind this table is on the benchmarks page, with the full settings.