SDXL base + refiner workflow
Stable Diffusion XL ships as two models working together: a base that establishes the composition and a refiner that polishes the final stretch of the denoising process. Splitting the work this way can sharpen faces, hair, and fine textures — provided the handoff point is set sensibly. This guide turns your total step count and switch point into the exact per-model settings.
How it works
In the recommended “ensemble of experts” setup, both models share a single latent. The base denoises from full noise up to a switch step, then the refiner continues from there to the end. The switch is usually expressed as a fraction — around 0.8 means the base does the first 80% of steps and the refiner the final 20%. Because the refiner only handles the low-noise tail, its job is detail, not structure: too large a share and it has nothing left to refine and can soften the image. This tool splits your total steps at your chosen point and reports the base steps, refiner steps, and the refiner’s effective denoise.
Why two models instead of one?
Standard diffusion models learn to denoise across the entire noise schedule — from completely randomised pixels down to the final image. SDXL’s designers observed that different parts of this process benefit from different model capacities: the early high-noise steps (where the base runs) establish large-scale composition and content, while the later low-noise steps (where the refiner runs) are almost entirely about sharpening fine detail.
Training a single model to do both jobs perfectly requires it to do very different things depending on the noise level. Splitting into two specialised models lets each be trained to excel at its specific part of the schedule. The result is better detail recovery at the end of generation than a single model of similar total size.
Choosing the switch point
The switch point (expressed as a fraction of total steps) is the most important tuning parameter:
| Switch fraction | Base steps (at 40 total) | Refiner steps | Effect |
|---|---|---|---|
| 0.7 | 28 | 12 | More refiner influence, risk of over-softening |
| 0.8 | 32 | 8 | Standard starting point |
| 0.85 | 34 | 6 | Less refiner, good for abstract or stylised art |
| 0.9 | 36 | 4 | Minimal refiner pass, mostly for detail sharpening on faces |
Higher switch fractions give the base model more of the process and the refiner less. If the refiner is softening or changing the composition, increase the switch fraction to give it fewer steps.
Ensemble vs img2img refining
There are two ways to use the refiner:
Ensemble of experts (recommended). The base and refiner share the same latent in one continuous pass. This is what this tool calculates — it is the cleanest and most controllable approach.
Img2img pass. Generate a full image with the base alone, then run it through the refiner as an img2img operation at a low denoise strength (typically 0.15–0.3). This approach lets you inspect the base output before refining, but adds a full second generation pass. The denoise strength for this approach is independent of the step count used in the initial generation.
Tips and trade-offs
- Start at an 80/20 split. It’s the most reliable balance of detail and speed for the original SDXL base + refiner pair.
- Test base-only first. Many community checkpoints already bake in the refiner’s strengths — if base-only looks great, skip the second model and save time.
- The refiner adds latency. Two models mean two loads and two passes; on limited VRAM the speed cost can outweigh the detail gain.
- Faces and hair benefit most. Reach for the refiner on portraits and detailed close-ups, less so for flat or stylized art.