Hires.fix is a two-pass Stable Diffusion workflow. The model first generates at a low base resolution it was trained on, then upscales and runs a second img2img pass to add detail. It produces sharper large images than generating at high resolution directly.

Why not just generate at the full target resolution?

Models are trained around a native resolution (512px for SD 1.5, 1024px for SDXL). Generating far above that causes duplicated limbs, repeated patterns and warped anatomy. Hires.fix keeps the first pass at the native size to avoid this.

What base resolution should SD 1.5 use?

SD 1.5 works best with a base around 512–768px on the long edge. The calculator divides your target by the upscale factor; if the result drifts far from that range, lower the factor or pick a different target.

How accurate is the VRAM estimate?

It is a rough guide based on pixel count and model. Real usage depends on batch size, sampler, attention optimisations like xformers, and whether you use tiled VAE. Treat it as a sanity check, not a guarantee.

SD Hires.fix Resolution Calculator

Get the base resolution for hires.fix right

Hires.fix gives Stable Diffusion images their crisp, high-resolution look — but only if the base resolution of the first pass lands in the model’s trained sweet spot. Set it too high and you get duplicated heads and warped bodies; too low and the upscale looks soft. This calculator takes the final size you want and an upscale factor and tells you exactly what base resolution to generate at, plus a VRAM sanity check.

How the hires.fix math works

Hires.fix is a two-pass pipeline:

Base pass — generate at a resolution close to the model’s native size (≈512px for SD 1.5, ≈1024px for SDXL).
Upscale + refine — multiply by the upscale factor and run a short img2img pass to paint in detail.

So the relationship is simply:

base_width  = target_width  / upscale_factor
base_height = target_height / upscale_factor

The trick is choosing an upscale factor that lands the base size near the model’s native resolution. If you want 2048×2048 from SDXL, a 2× factor gives a 1024×1024 base — perfect. The same target on SD 1.5 would want a ~3× factor to keep the base near 680px.

Tips to avoid out-of-memory and artifacts

Stay near native. Keep the base long edge within roughly ±50% of the model’s native resolution.
Watch VRAM on the second pass. The upscaled pass operates at full target resolution — that is where most cards run out of memory. Enable tiled VAE if the estimate is tight.
Use a low hires denoising. 0.3–0.5 adds detail without changing the composition the base pass established.
Multiples of 8. Both base and target dimensions should be divisible by 8; the calculator rounds for you.