GigaHires Easy

Feed GigaHires Easy your first-pass latent, pick a scale, a quality, and a detail level, and it runs a complete hires fix: upscale the latent, then run a second-pass refinement sample. Every knob the other nodes expose is hidden behind three dropdowns. It's the "just make it bigger and better" button, and for a first try that's exactly what you want.

What it's for (and what it isn't)

The pack README is blunt about this one: it's a "simplified wrapper kept around for quick experiments, not the main UX direction." The project's recommended path is the helper-node workflow - keep the normal pass, the upscale, and the refine step visibly separate. Easy is the opposite: one box, everything inside. Use it to get a feel for what a good hires fix does, then graduate to GigaHires Latent Upscale / GigaHires Image Upscale plus GigaHires Refine Pass the moment you want control.

How the presets map

Behind the scenes it translates quality and detail into hardcoded settings, then runs the same two-pass pipeline the big coordinator node uses:

• detail sets the refinement strength: subtle → denoise 0.22 / cfg 6.5, balanced → 0.35 / 7.0, strong → 0.5 / 7.5.
• quality sets the budget: fast → 8 steps, euler + normal scheduler; balanced → 12 steps, dpmpp_2m + karras; high → 16 steps, dpmpp_2m + karras, and it flips the VAE into tiled mode.

Everything else is fixed for you: mode picks the upscale branch (latent by default, or upscale_model if you want a learned upscaler), sizing is always scale with your scale_by, and the latent interpolation is always antialiased. That last one is a good default, because antialiased is the mode the author's testing landed on.

The inputs that matter

• scale_by - the upscale factor (2 is the sweet spot per the broader upscaling playbook; 1.5–2x keeps composition safe).
• detail - the one you'll actually change. strong at 0.5 denoise starts drifting the composition; subtle barely moves anything.
• mode - upscale_model only makes sense if you've also selected a model, see below.

Outputs: refined_latent and refined_image are the results (either can feed a save or continue the graph), and debug_info is the usual JSON dump - this one adds easy_mode: true and the pass-2 image shape so you can confirm what the presets actually chose.

Where people get burned

Same trap as the rest of the pack: upscale_model_name defaults to "None". If you set mode = upscale_model without picking a model or wiring the optional upscale_model input, you get a ValueError on run. Pick a name or stay on latent.

Also expect high quality to be noticeably slower - tiled VAE is a VRAM safety net, not a speed hack. And there's no way here to target an exact pixel size, override the second-pass conditioning, or change the sampler. That's the point of the node, but it means the ceiling is the presets.

Install

Manager → search "comfyui-hires", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Enferlain/comfyui-hires

Restart ComfyUI. No requirements.txt, no pip installs - the only real prerequisite is a recent ComfyUI, because the pack is built on the newer native extension API and silently fails to load on old builds.

Bottom line

It's a perfectly good entry point and a fine way to A/B a hires factor without touching ten widgets. Just don't treat it as the pack's flagship - the author doesn't, and the moment you want to know why an image is blocky or slow, you'll be reaching for the helper nodes and this one's debug_info will have pointed you there.