Steps, CFG and samplers
Switching models switches to that architecture’s recommended values
Section titled “Switching models switches to that architecture’s recommended values”Architectures want very different steps, CFG and samplers (SDXL needs 28 steps; FLUX.2 klein is done in 4). When you switch to a model of a different architecture, these change together:
- If you’ve used that architecture before, you get your last settings for it.
- The first time, you get the official recommendation:
| Architecture | Steps | CFG | Sampler |
|---|---|---|---|
| SDXL, SD 1.5 | 28 | 5 | Euler a |
| Anima | 30 | 4 | ER-SDE |
| Z-Image | 28 | 4 | Euler |
| Krea 2 (Turbo) | 8 | 1 | Euler |
| Qwen Image 2.1 | 40 | 1 | Euler |
| FLUX.2 klein (distilled) | 4 | 1 | Euler |
| FLUX.1 dev | 20 | 3.5 | Euler |
| FLUX.1 schnell | 4 | 1 | Euler |
| FLUX.1 Kontext | 28 | 2.5 | Euler |
For FLUX.1 dev and Kontext, CFG is the official guidance value: each step is computed once and the negative prompt isn’t used.
- “Common sizes” lists aspect ratios at about one megapixel (what SDXL was trained on), plus the slightly larger sizes Anima likes.
- “Swap” exchanges width and height.
- Save sizes you use often with “Save current size” and pick them later under “My sizes”.
- Width and height are rounded to a multiple of 16.
- Random: a different one for every image.
- Fixed: the whole batch shares one seed — use it for comparisons (for example several LoRAs, so the only difference is the LoRA).
- “Use the last seed”: for small tweaks to an image you like.
Press “More” to open these. Most of the time you can leave them alone:
| Option | When to change it |
|---|---|
| VRAM | Speed: fastest, uses the most graphics memory. Balanced: right most of the time; saves a lot and is only a little slower. Low: only needed at 8 GB or less; uses the least but is several times slower |
| VAE tiled decode | Auto tiles above 1280×1280. Pick “Always” when graphics memory is short |
| Staged decode (SDXL only) | Moves the model off the graphics card before turning the result into an image. Saves memory, same image, a second or two slower |
| Conditioning | Leave on Auto. a1111 supports weight syntax and long prompts; native is the plain diffusers behaviour |
| CLIP skip | The minimum for SDXL is 2 |
| shift | Only for Anima, Z-Image, Krea 2, Qwen and FLUX, remembered per architecture. Higher values change the composition more |