Running FLUX on an 8 GB GPU
FLUX.2 klein 4B fits on an 8 GB card as FP8, at 6.2 GB of VRAM, and makes and edits images in 4 steps. FLUX.3 klein doesn't fit (12.9 GB at peak), but Latent streams it through an 8 GB card in about 60 seconds a frame if your RAM can hold it, or about 111 seconds from disk.
The two FLUX models side by side
| FLUX.2 klein 4B | FLUX.3 klein | |
|---|---|---|
| Does | generate and edit, 4 steps | generate, and edit with up to 10 reference photos, 16 steps |
| VRAM at peak | 9.8 GB (6.2 GB as FP8) | 12.9 GB |
| On a 16 GB GPU | about 12 s per image or edit | about 17 s per image, 33 s per edit |
| Streamed from RAM | 23 s | 60 s |
| Streamed from disk | 25 s, 8.2 GB RAM | 111 s, 11.1 GB RAM |
| On the CPU | 177 s, 18.6 GB RAM | not measured |
| Download | 16 GB | 20 GB |
| License | Apache 2.0 | FLUX Kommunity License: commercial use free under US$5M revenue |
Which one to use on 8 GB
- For speed, FLUX.2 klein 4B. As FP8 it runs entirely on the card. Even when streamed it stays near 25 seconds, because the model is small.
- For combining several photos, FLUX.3 klein. It takes up to 10 reference photos and keeps what you didn't ask to change. Expect about a minute per frame when streamed.
With Settings › Run models on set to Automatic, Latent picks the fastest variant that fits your VRAM and RAM by itself. Hover over a model in the top bar to see its footprint before loading it.
FLUX.3 klein is gated on Hugging Face: free, with instant approval. Make a free account, open the model's page and click Agree and access repository, create a Read token, and paste it into Latent's Settings. The token stays on your computer and is only sent to Hugging Face.
Editing a photo with FLUX
- Drop, paste (Ctrl V) or import a photo onto the window.
- Describe the change in the prompt and press Ctrl Enter.
- Press E on a result to use it as the source for the next edit, without saving and importing it again.