Methods
What we measure
We measure how long a consumer GPU needs to make one image with an open model, and we keep the pictures so you can judge the quality yourself. Speed tables are on Results; the pictures are in the comparison viewer and on each model page.
Where the machines come from
We rent time on gaming PCs owned by individuals. Each machine runs a container inside a Linux virtual machine on Windows, with ComfyUI as the image engine. We do not identify the owners. A "machine" in our tables is one rented node, counted once.
How one number is made
- Same inputs. Every model gets the same prompts (version v1 and a more detailed v2) and the same seeds (300 and 301 for the gallery pictures; the portrait prompt for timing). Size is 1024×1024 unless a page says otherwise.
- Model settings. We use the settings the model's authors recommend (steps, sampler, CFG). They differ between models on purpose, so a speed table compares each model at its own recommended settings, not at identical ones. Every model page lists its settings.
- Warm runs only. The first image on a machine includes loading the model and is reported separately as "start to first image". The speed number is the median of 5 warm runs on that machine, using the server-side total time of the ComfyUI API. We change the seed on every run, because ComfyUI caches an identical request and returns it almost instantly (see the 0.1-second trap).
- Several machines. The figure for a GPU is the median of its machine medians. We show the range across machines. A figure from fewer than 5 machines is marked * and is reference only.
- VRAM and cost. Peak VRAM is the value reported by nvidia-smi and includes memory PyTorch reserves, so it is not the minimum needed. Cost per image is the rental hourly price at the time multiplied by the server time, so it is rough.
Raw GPU numbers
We also run a short PyTorch test on each machine (matrix speed and memory bandwidth). It is our own script and the machines vary a lot, so read it with the note that explains how our first answer did not survive more machines.
Pictures
Pictures are made by us with the settings on each model page. Every published picture carries our watermark and passes a check for readable names, signatures and unsuitable content before it appears; where a model reproduced a real artist's signature, we blurred that corner and said so in the note. Pictures are for all ages. Cloud services such as GPT Image are shown as references, with no seed match, and are never in the speed tables.
Votes
The thumbs-up counts show personal preference, are not verified, and say nothing about accuracy. A one-way, daily-changing hash of the connection limits repeat votes; we do not store the address itself.
What this site cannot tell you
- Machines differ: the same GPU varied by a few percent to about 20 percent between rented machines, and some raw numbers varied by 50 percent.
- One prompt is used for timing, so heavy prompts, other sizes or other software can change the ranking.
- We do not score image quality. Two or three seeds per prompt are anecdotes, not statistics.
- We can be wrong. When we find a mistake we fix it and say so in the note: see the correction in the MOSS-Image note.