Comparison table
| SD 1.5 | SDXL 1.0 | SD 3 Medium | SD 3.5 Large | |
|---|---|---|---|---|
| Release | Oct 2022 | Jul 2023 | Jun 2024 | Oct 2024 |
| Parameters | 0.9 B | 2.6 B (+ refiner) | 2 B | 8 B |
| Native size | 512×512 | 1024×1024 | 1024×1024 | 1024×1024 |
| VRAM (fp16) | 4–6 GB | 8 GB | 10 GB | 19 GB (11 GB in fp8) |
| Speed (our 12 GB GPU) | ~4 s | ~20 s | — | — |
| Prompt understanding | keywords | keywords, better composition | sentences, several subjects | best |
| Text in images | no | rarely | yes, short | yes |
| Artist names | strong | strong | weaker | weaker |
| Community models | very many | the most, with Pony and Illustrious | few | few |
| Licence | OpenRAIL-M | OpenRAIL++ | Community licence (free under $1M revenue) | |
| Try on this site | SD 1.5 | SDXL | SD 3 | — |
Stable Diffusion 1.5: the workhorse
Small, fast, endlessly customised. Photorealistic and anime fine-tunes, ControlNet models, LoRAs and embeddings exist for everything. Its weaknesses are the 512 px base resolution, hands, and a literal reading of prompts. It remains the right choice for a 6 GB card, for AnimateDiff animation and for ControlNet-heavy work. The members' txt2img page on this site runs a custom SD 1.5 model.
SDXL 1.0: detail and style
SDXL has four times the pixels, a second text encoder and a refiner for the last steps. Faces, hands and composition improved a lot, and artist names have a strong, predictable effect, which is why our 1,731-artist comparison was generated with SDXL. It needs 8 GB and is several times slower than 1.5: about 20 s against 4 s per image on our GPU. Distilled versions (SDXL Turbo, LCM LoRA) need only one to eight steps instead of 25 to 30, at some quality cost.
Stable Diffusion 3 and 3.5: understanding and text
A new architecture (diffusion transformer with three text encoders) that follows sentences instead of keyword lists: most of the time "a red cube on a blue sphere, to the left of a green cone" comes out right and short text in quotes is readable. 3 Medium had anatomy problems that 3.5 reduced; 3.5 Large is the most capable of the family, 3.5 Medium the balanced one. Artist names and old negative prompt habits matter less. Try Stable Diffusion 3 Medium here.
Which one should you pick?
- Learning, 6 GB card, ControlNet, AnimateDiff: SD 1.5.
- Illustration, portraits, artist styles, 8–12 GB: SDXL.
- Posters with text, complex scenes, 12–16 GB+: SD 3.5 Medium or Large.
- No GPU: the online tools on this site, no account needed.