Local AI image models compared
Fastest on a 16 GB GPU: SANA-Sprint 1.6B at about 7 seconds per 1024 × 1024 image. Smallest: SANA-Sprint at 5.3 GB of VRAM. For editing photos: FLUX.3 klein, FLUX.2 klein 4B or Qwen-Image 2.1 Turbo. Free for commercial use: the Apache 2.0 models, plus FLUX.3 klein under US$5M revenue.
What each model does
| Model | Does | Steps | License |
|---|---|---|---|
| Z-Image Turbo | generate | 9 | Apache 2.0 |
| FLUX.3 klein | generate, edit with up to 10 photos | 16 | FLUX Kommunity (commercial use free under US$5M revenue) |
| FLUX.2 klein 4B | generate and edit | 4 | Apache 2.0 |
| Qwen-Image 2.1 Turbo | generate and edit, strong on text and layouts | 8 | Qwen research license, non-commercial |
| SANA-Sprint 1.6B | generate | 2 | Apache 2.0, Gemma terms for its text encoder |
| Kromoss | generate (a 5B MOSS-Image fine-tune) | 8 | Apache 2.0 |
Memory
| Model | VRAM at peak | RAM streaming from disk | RAM on the CPU |
|---|---|---|---|
| Z-Image Turbo | 14.2 GB | 18.9 GB | about 23 GB |
| Kromoss | 13.7 GB (9.0 GB as FP8) | 16.1 GB | about 29.6 GB |
| FLUX.3 klein | 12.9 GB | 11.1 GB | about 22.6 GB |
| Qwen-Image 2.1 Turbo | 10.5 GB | 17.1 GB | about 34.9 GB |
| FLUX.2 klein 4B | 9.8 GB (6.2 GB as FP8) | 8.2 GB | 18.6 GB |
| SANA-Sprint 1.6B | 5.3 GB | 6.4 GB | 11.9 GB |
Speed
| Model | On the GPU | Streamed from RAM | Streamed from disk | On the CPU |
|---|---|---|---|---|
| SANA-Sprint 1.6B | 7 s | 12 s | 12 s | 37 s |
| FLUX.2 klein 4B | 12 s | 23 s | 25 s | 177 s |
| Kromoss | 12 s | not offered | 64 s | not measured |
| Z-Image Turbo | 13 s | 57 s | 108 s | not measured |
| FLUX.3 klein | 17 s | 60 s | 111 s | not measured |
| Qwen-Image 2.1 Turbo | 20 s | not offered | 74 s | not measured |
Edits take longer than new images: about 33 s for FLUX.3 klein and 39 s for Qwen-Image 2.1 Turbo on the GPU. Qwen-Image 2.1 Turbo also uses 20 GB of RAM.
Picking one
- 16 GB card: start with Z-Image Turbo. Read how much VRAM Z-Image Turbo needs.
- 8 GB card: FLUX.2 klein 4B or SANA-Sprint 1.6B. Read running FLUX on 8 GB.
- No GPU: SANA-Sprint 1.6B. Read generating images without a GPU.
- Text in images, posters, layouts: Qwen-Image 2.1 Turbo, for non-commercial work.
Latent also offers Qwen-Image 2.1 Uncensored, which has no content filter and is non-commercial. It warns you and asks before downloading it.
Each model's license applies to what you make with it. Check it before commercial use; this table is a summary, not legal advice.