# FLUX.2 Klein 4B on a 3090, 4090, 5090 and A5000: the same setup, four rented cards

> We built a FLUX.2 Klein 4B text-to-image setup once, then restored it on four different cards. How long each took, what the pictures looked like next to the original, and what to take from it if you rent by the hour.

Published: 2026-10-12 · Field reports
Canonical page: https://renest.ai/blog/flux2-klein-on-four-gpus

**Short answer:** all four cards brought the setup back and rendered a picture. The 5090 was done in just over a minute and a half; the 4090 took three and a half minutes. The render itself took 6 to 12 seconds on every card, so the card wasn't what made the difference. Next to the original picture, the pictures from the cards we opened show the same scene and composition. On the 4090 and 5090 the pictures differ from the original in small details; on the 3090 we couldn't tell the two apart.

## The setup

FLUX.2 Klein 4B is a small, fast text-to-image model, so it is a common first thing to rent a GPU for. Ours was a plain ComfyUI text-to-image workflow: the fp8 diffusion model, the Qwen3 4B text encoder and the VAE, 12.5 GB in all. We built it on a RunPod A40: 33 seconds to render the first picture, 75 seconds to pack the whole thing into a nest.

Then we rented four other machines, one after another, and on each pasted one restore command from our drive.

## What each card did

| Card | Download | Render | Start to finished picture | Similarity to the A40 picture |
|---|---|---|---|---|
| RTX 5090 | 33 s | 6 s | 1 min 37 s | 0.950 |
| RTX A5000 | 107 s | 9 s | 2 min 49 s | 0.995 |
| RTX 3090 | 98 s | 12 s | 2 min 59 s | 0.994 |
| RTX 4090 | 101 s | 9 s | 3 min 30 s | 0.953 |

"Start to finished picture" is the whole restore, from pasting the command to the test render finishing. Render is the part you'd think the card decides: it ranged from 6 to 12 seconds. The download was the longest single stage on three of the four rows, 98 to 107 seconds, and 33 seconds on the 5090. The 4090 had the slowest total even though it rendered in 9 seconds: the stages between download and render (putting files in place, installing dependencies, starting the app) added up to about 100 seconds there, against 53 to 69 seconds on the other three. We didn't dig into why.

The cards ran the tool in different versions: the A5000 on 0.1.20, the others on 0.1.18. All four brought the files back, checked, and rendered.

## How close were the pictures?

We compare each restored picture with the one the A40 made from the same prompt and seed, using structural similarity (1.0 is the same pixels). The A5000 and 3090 came out at 0.995 and 0.994. The 4090 and 5090 came out at 0.953 and 0.950.

Here are the three cards whose pictures we compared, each next to the A40 original: original on the left, restored picture in the middle. The right-hand tile is the difference between the two, with every difference tripled so that it can be seen at all; black means the same pixels.

![Three rows, one per card (RTX 3090, RTX 4090, RTX 5090). Each row shows the A40 original, the restored picture, and the difference between them tripled. All six pictures show the same hedgehog in a party hat on a confetti-covered floor; the difference tiles are mostly black with fine light lines along edges.](/blog/flux2-klein-three-cards.jpg)

Side by side, all three show the same hedgehog in the same party hat, in the same pose and the same room. None is pixel-identical (the similarity numbers are in the table above), but even with the differences tripled the difference tile is mostly dark: fine light lines along the hat and the spines and a few bright spots on the confetti, and on the 3090 almost nothing. On the 3090 pair we couldn't tell the two pictures apart; on the other two you have to look closely to see any difference.

On a different GPU model that number isn't a pass or a fail, so we list it as a fact. Floating-point maths comes out slightly differently on different chips, and a diffusion model amplifies small differences. If you need the same pixels, go back to the GPU model you made the picture on. If you want a working setup on a new card, any of these is fine.

## What we'd tell someone renting by the hour

- **Don't pick by render time alone.** Render was 6 to 12 seconds on every card. Everything else around it was 90 seconds or more, and it varied more from machine to machine than the render did.
- **Keep the setup next to the cards, not on one machine.** Each of these was a fresh machine with nothing on it. The restore command was the same on all four.
- **Don't expect the same pixels on a different model of card.** The similarity number was 0.995 and 0.994 on two of the cards and 0.953 and 0.950 on the other two. Where we opened the pictures, it was the same scene with small differences in detail on the lower-scoring ones.

## What this doesn't say

Each row is a single run. These are RunPod machines in whatever region had stock, so the download column reflects the connection we happened to get. Times on a different day will differ. The similarity numbers are measurements of one prompt and one seed.

If you want to repeat it, the starter nest is on your Renest drive for free (see [Try a starter nest](/docs/starter-nests)).