SD1.5 NPU High Resolution

WARNING

Quantized high-resolution (>768×768) SD1.5 models often produce images with poor overall layout — extra limbs, distorted proportions, repeated subjects. This is a known limitation of W8A16 static quantization at large resolutions, not a bug in your setup.

Recommended Workflow (Highres.fix)

The reliable way to get high-resolution images on the NPU is a two-stage process, equivalent to AUTOMATIC1111's Highres.fix:

1. Generate at 512×512 with the base SD1.5 NPU model. The composition will be coherent.
2. (Optional) Upscale the result with one of the built-in upscalers if you need a sharper input.
3. Run img2img at the target resolution. Use a denoise strength around 0.8 so the output keeps the original layout but adds high-resolution detail.

Step 1: txt2img @ 512×512   →  good layout, low detail
Step 2: img2img @ target     →  good layout + high detail
        denoise_strength ≈ 0.8

In-app steps for stage 2

1. From the result screen of your 512×512 generation, use 1-click img2img to send the image straight into the img2img tab (see History → 1-Click img2img).
2. Tap the resolution selector on the model run screen and pick the target resolution (e.g. 768×1024). This is the manual step that applies the resolution patch — the UNet reloads for 1–2 seconds.
3. Set denoise strength to ~0.8 and generate.

Available Patch Resolutions

The following resolutions are available as patches on top of the 512×512 base (requires Snapdragon 8 Gen 1 or above):

Resolution	Aspect Ratio
512 × 768	2:3
768 × 512	3:2
768 × 1024	3:4
1024 × 768	4:3

You can switch between resolutions directly from the model run screen without exiting.

Why Not Just Generate at 1024x1024 Directly?

Direct high-resolution generation goes through the same quantized weights but without the structural prior of a small initial latent. The model has to infer the entire composition at high resolution, which is exactly the case where quantization noise hurts the most.

The two-stage workflow lets the low-resolution pass do the structural work and uses the high-resolution pass only for refinement, where quantization is much more forgiving.

Tips

• A denoise strength below 0.7 often loses high-frequency detail; above 0.85 can shift the composition.
• For SDXL NPU models, this workflow is unnecessary — SDXL runs at 1024×1024 internally and supports arbitrary aspect ratios directly. See SDXL Aspect Ratio.
• The upscaler step before img2img is optional but helps when going from 512 to a larger resolution.
• Non-flagship chips (_min tier) do not support high-resolution patches.
• Changing to a resolution with a different aspect ratio will clear the current input image.