Same seed, same picture? 22 machines gave 9 versions of one SDXL image
If you give an image model the same prompt, the same settings and the same seed, do you get the same picture on another computer? We had the files to check it without new runs: on 2026-10-23 we made the same SDXL portrait with seeds 100 to 105 on 22 rented machines with five different GeForce cards, for our speed table. We compared every picture pixel by pixel, after decoding, so the text stored inside the PNG files does not count.
What we compared
Every machine ran the same container image (ComfyUI 0.35.0, PyTorch 2.13, CUDA 13.0) and the same workflow: SDXL 1.0 base, 1024×1024, 30 steps, the euler sampler, CFG 7, one portrait prompt. Only the machine changed. Every picture was really computed: we change the seed between runs because ComfyUI returns a stored result when a job is repeated exactly (the 0.1-second trap). We also had a smaller set made the same way with FLUX.1 schnell (5 machines) and FLUX.2 klein (4 machines).
What we found
To the eye, every machine made the same picture. In the pixels, they did not. For each SDXL seed, the 22 machines made 9 different versions:
| Card | Machines | Versions among them |
|---|---|---|
| RTX 3060 12GB | 6 | 2 (4 machines and 2 machines) |
| RTX 4060 8GB | 4 | 2 (3 and 1) |
| RTX 4070 12GB | 4 | 1 |
| RTX 5060 8GB | 4 | 2 (2 and 2) |
| RTX 5070 12GB | 4 | 2 (3 and 1) |
No version was shared between two kinds of card. And the groups did not change from seed to seed: machines that matched on seed 100 also matched on seeds 101 to 105, and machines that differed on one seed differed on all six.
How large are the differences? Compared with the largest group, 59 to 78 % of the pixels differed by at least one level (out of 255 per colour), but the average difference was only 0.6 to 1.7 levels. In a few places a single pixel was off by more than 100 levels, which is still too small an area to see at normal size.
The FLUX models behaved the same way. FLUX.1 schnell made 3 versions on 5 machines (one per kind of card: two RTX 3060 matched, two RTX 5070 matched), with 86 to 92 % of pixels differing and an average of 2.0 to 3.4 levels. FLUX.2 klein made 3 versions on 4 machines: the two RTX 5070 matched, the two RTX 4070 differed from each other and from the RTX 5070.
Same card, later
We also have one case of the same card on different days. On 2026-11-02 we made 30 images on an RTX 5070 (SDXL, and MOSS-Image SFT with 25 and with 8 steps, on our standard prompts) that we had first made between 2026-10-23 and 2026-10-26, also on rented RTX 5070 cards. All 30 matched pixel for pixel. We did not record which machines made the earlier images, so we cannot say whether they were different machines.
What we do not know
We did not find out what decides which version a machine makes. It is not only the kind of card, because cards of the same kind split into groups. It is also not the driver version alone: one RTX 3060 group has machines with drivers 610.88, 617.42 and 621.01, and the two RTX 5070 machines that matched on FLUX.2 klein had drivers 591.86 and 621.01. We did not test other possible causes.
What this means for you
- A shared seed and settings will give you the same picture, as far as anyone can see, on a different card. Do not expect the files to be identical.
- If you compare images by file checksum, for example to test that an update did not change your output, expect a mismatch when the machine changes, even with the same kind of card. A check that allows small differences works better.
- This does not affect the pictures in our Compare pages: they come from different machines, and a different machine changes the pixels this little.
Limits: one prompt and six seeds for SDXL; one machine of each kind for some FLUX versions; rented machines only. We kept the comparison script and the numbers for every seed and machine.