UltraFix

UltraFix is a tiled-diffusion img2img repair pass for upscaled images. A plain upscaler (Real-ESRGAN / UltraSharp) enlarges an image but cannot invent real detail; UltraFix runs the diffusion model again over the enlarged image in overlapping tiles to restore fine texture and detail at high resolution, without re-staging the whole composition.

Only on SDXL DMD2 models

UltraFix is available only on SDXL DMD2 models (NPU path), because it relies on the CFG 1 / few-step setup those models use. It also requires that img2img is enabled for the model — with img2img turned off the backend has no VAE encoder and the button is hidden.

Just as important: only use it on anime models. On realistic models it tends to duplicate subjects (see Usage Tips).

Where to Find It

After you upscale a generation, the result is larger than the model's normal generation size. When the upscaled result's shortest edge is at least the model's generation size, an UltraFix button appears:

• on the generation result page
• in the history preview overlay (same gating as the result page)

The repaired image is saved as a new ULTRAFIX history entry (JPEG), and you can run UltraFix on it again.

How It Works

UltraFix is based on the partial-inversion + noise-injection approach from the PixelRush paper, adapted to run as fixed-size NPU tiles.

The enlarged image is split into overlapping tiles. The VAE encode, every UNet denoising step, and the final VAE decode all run the fixed-size QNN graphs over those tiles; per-step noise predictions are blended across the tile overlaps before a single image-wide scheduler step, so there are no visible tile seams.

Instead of destroying the image with random noise (ordinary img2img), UltraFix:

• Partially inverts the clean image back up the diffusion trajectory (deterministic DDIM inversion), so the image's own structure lives inside the starting noise.
• Locks low-frequency structure to the base image early in the loop, so tiles keep the original composition instead of each restaging the prompt subject.
• Injects a small amount of fresh noise later in the loop (PixelRush Eq. 4), which restores high-frequency detail that the near-reconstructive inversion would otherwise leave flat.

You do not need to understand any of this to use it — but it explains why UltraFix adds detail rather than redrawing the picture.

Controls

UltraFix has its own Steps and Denoise sliders, separate from the main generation parameters and persisted per model:

• Steps — 1–20.
• Denoise — expressed as a step count (0 … min(10, steps)) rather than a raw strength. A higher denoise-step count lets UltraFix change more of the image; a lower count stays closer to the input.

A Restore Defaults button resets these to 10 steps / 4 denoise steps.

Everything else — prompt, negative prompt, CFG, scheduler, and seed — is taken from the prompt page, so set those up there before running UltraFix.

Usage Tips

• 10 steps / 4 denoise steps are the recommended defaults. The main dial is the denoise-steps / steps ratio — keep it ≤ 0.5.
• Because each tile only sees a crop of the image, the model can re-interpret that crop as a full scene and paint an extra subject into it (an extra face, a second character, a duplicated object). If you see this, increase steps or lower denoise steps — e.g. 12 / 4 or 10 / 2. That reduces how much UltraFix changes per pass, but you can simply run UltraFix again on the result to keep building up detail.
• Only use UltraFix on anime models. On realistic models the multi-subject problem above is much more likely, so it is not recommended there.
• Treat it as a finishing pass on an image you already like, not a way to rescue a bad generation.
• It is slow and hot: a full UNet runs over every tile for every step at high resolution. The confirmation dialog lists the parameters and warns about runtime and device heat before you start. Keep the app in the foreground while it runs.
• Generation, upscaling, and live previews are disabled while UltraFix is preparing and running.

Related

• Upscalers — the plain 4× enlargement step that comes before UltraFix.
• DMD2 Models — the CFG 1 / few-step models UltraFix is tuned for.