GigaHires V1

One node that takes your first-pass latent and runs the entire hires fix: upscale it via the latent path or a learned upscaler, then run a second refinement pass. It's the pack's "most knobs exposed" coordinator - and it's the direction the README has since politely demoted. The recommended workflow in this pack is the helper-node layout, keeping the normal pass, upscale, and refine visibly separate. V1 is the everything-in-one-box version that predates that decision. It works, and it's a good way to see the whole pipeline at a glance; just know what you're trading away.

What it does

Internally it's one function doing three visible stages, and the outputs mirror them:

1. Branch upscale - branch_mode = latent interpolates the latent up to your target (latent_mode controls the filter); branch_mode = upscale_model decodes, runs a learned upscaler, re-encodes. sizing_mode / scale_by / target_width / target_height work exactly like the helper nodes.
2. Refine pass - a second KSampler at your denoise (default 0.35) with its own steps (12), cfg (7), sampler_name, scheduler, and the vae_mode / vae_tile_size / vae_overlap trio for big decodes.
3. Decode - to the refined_image.

That's why there are five outputs and it can feel like a lot: pass2_latent / pass2_image are the state between upscale and refine, and refined_latent / refined_image are the finished thing. Save the refined ones. debug_info is the JSON report - branch used, model name, base vs resolved size, latent shapes, per-stage timings.

The inputs worth knowing about

• branch_mode - latent is the default and the cheap path; upscale_model is the quality path but requires a model (see the trap below).
• positive_pass2 / negative_pass2 (both optional) - the one thing the helper nodes don't give you in one box. Wire these to override the conditioning for the second pass, e.g. a detail-focused prompt, while the first pass keeps your main prompt. Leave them empty and pass 2 reuses the first-pass conditioning.
• sampler_name / scheduler - and here's a genuine quirk: this node defaults to euler + normal, while the pack's own GigaHires Refine Pass defaults to dpmpp_2m + karras, which the author's notes name as the quality baseline. You can change it, and you probably should.

Where people get burned

• The upscale_model ValueError. Set branch_mode = upscale_model without picking an upscale_model_name or wiring the optional upscale_model input and the run fails on purpose. Pick a model first.
• Reading the wrong outputs. pass2_* is mid-pipeline; if you're saving pass2_image and wondering why it's rough, that's the not-yet-refined one.
• Latent branch blockiness. The author's testing found pure latent upscale can stay blocky even after refinement at low denoise. If you hit it, branch_mode = upscale_model or a larger target is the move, not more steps.
• All-in-one opacity. The real cost of this node is that you can't inspect each stage's debug_info separately - which is exactly why the author built the helper nodes afterward.

Install

Manager → search "comfyui-hires", or:

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

Restart ComfyUI. No pip dependencies - the pack relies on ComfyUI internals and the native extension API, so keep ComfyUI current or the nodes won't register at all.

Bottom line

V1 is a fine way to wrap your head around what a two-pass hires fix does before you care about the stages individually. Once you do care - and with this pack, that's the intended journey - the helper nodes give you the same pipeline with visibility. V1 is the map; the helpers are the territory.