HiResFix Parameters

HiResFix Parameters is the node that controls Searge's high-resolution pass - the second generation step that pushes your image past native resolution while adding coherent detail instead of just stretching pixels. If you want bigger, sharper images out of the Searge workflow, this is the panel.

It's part of SeargeSDXL, the base-plus-refiner SDXL workflow pack Searge shipped at SDXL 1.0's launch in 2023. It sits in Searge/_deprecated_/UI/Inputs; the v4.x rewrite moved to a data-stream design, but hi-res fix itself is a technique you'll use everywhere, not just here.

How it works, and why hi-res fix exists

SDXL generates natively around 1024x1024. Ask it directly for something much bigger and you get duplicated limbs and repeated patterns, because the model wasn't trained at that size. Hi-res fix dodges that: generate at native resolution, upscale the latent, then run a second sampling pass at low denoise so the model refines the upscaled image without redrawing the composition. The knowledge base calls this "essential for any generation above native resolution," and the key setting is the denoise - the community rule of thumb is to keep it low (roughly 0.3–0.5) so the second pass adds detail without changing the picture.

This node just collects the hi-res settings into the PARAMETER_INPUTS bundle; the actual second pass runs downstream.

The inputs that matter

• hires_fix (enum enabled/disabled) - the master switch.
• hrf_denoise (FLOAT, default 0.1) - how much the second pass is allowed to change. This is the one that makes or breaks the result; too high and the composition drifts, too low and it does nothing. The default is conservative - many people nudge it up.
• hrf_upscale_factor (FLOAT, default 1.5) - how much bigger. 1.5–2x is the sane range.
• hrf_steps (INT, default 0) - steps for the pass (0 means derive from the main step count).
• hrf_intensity (soft/hard), hrf_seed_offset (same/distinct), hrf_smoothness (FLOAT) - finer tuning; defaults are fine to start.
• inputs in/out (PARAMETER_INPUTS) - the chain wire.

How to install it

ComfyUI Manager: search SeargeSDXL, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/SeargeDP/SeargeSDXL.git, then restart. Manual installs need python -m pip install opencv-python in ComfyUI's Python environment first, per the README. The hi-res pass works best with a good upscale model - grab 4x-UltraSharp from the README's list along with the SDXL base/refiner checkpoints.

Common issues

The single biggest hi-res mistake is denoise too high on the second pass. Above ~0.5, the knowledge base warns, "the composition changes" - you'll get a different image, not a bigger version of the same one. If your hi-res result doesn't match the preview, drop hrf_denoise. If it looks identical to the input with no added detail, raise it a touch.

Note also that a bug in early v4.0 left the hi-res latent detailer doing nothing at all - the pack fixed that in v4.1, which is one more reason to run the latest node version against the latest workflow JSON. And the broad caveat: this is a deprecated node in a refiner-era workflow. Hi-res fix is still a great technique, but you don't need Searge to get it - most modern SDXL workflows include their own.