AUN KSampler PlusV3

AUN KSampler Plus (that's the AUNKSamplerPlusv3 class) is the pack's answer to one of the most common ComfyUI questions of all: "how do I do the hi-res fix without a spaghetti of nodes?" The hi-res fix - generate at native resolution, upscale, re-sample at low denoise to add coherent detail - is still the right move for generating above native resolution, and this node runs the whole chain in a single KSampler: base pass, latent upscale, second pass, optional pixel-space upscale, optional final refinement. Five images come out the other side, and you pick the one you wanted.

How it works

The flow is progressive, and worth internalizing because the outputs are named after its stages:

1. Base image - first pass, your normal generation (denoise 1.0 for txt2img).
2. Latent upscaled - the latent is upscaled (bicubic is the recommendation) by ratio, then re-sampled for a second pass.
3. Both upscaled - the decoded latent is pixel-upscaled, then re-sampled again mirroring the second pass's schedule. This is the "proper" hi-res-fix result.
4. Refined image - optional: re-encode and re-sample a chosen source with img_refine_steps / img_refine_denoise if you enabled image_upscale_refine.

The Upscaled type string output reports which one you ended up with - wire it into a filename builder so your saved file names itself honestly. You also get the final LATENT and each stage's image.

Inputs that matter

The full list is long; the ones a beginner actually tunes:

• steps_total and steps_first - total steps split between passes. steps_first is your base-pass steps.
• latent_upscale (on by default) and ratio (1.5–2.0 typical) - the core of the progressive feature.
• upscaling_denoise - second-pass denoise. The tooltip's guidance of 0.5–0.7 is in line with hi-res-fix practice, but remember the rule from the upscaling world: below 0.5 keeps composition tighter, higher changes more.
• start_step_second - -1 uses the denoise fraction, 0 restarts at step 0, steps_first continues where pass one left off.
• cfg vs cfg_latent_upscale - separate CFG for each pass, so you can drop CFG on the upscale leg.
• image_upscale / image_upscale_method / image_upscale_model / image_upscale_ratio - the pixel-space leg. Lanczos for photos, nearest for pixel art; pick an AI upscale model if you want it.
• verbose - prints per-pass logs and timings, which is exactly what you want while you're learning what each stage does.

Recommended pairing per the tooltips: DPM++ 2M with Karras, and AYS schedulers work well with the progressive schedule.

Installing it

Part of the AUN pack:

• ComfyUI Manager: search "AUN ComfyUI Nodes", install, restart.
• Manual: cd custom_nodes && git clone https://github.com/loz2754/AUN-ComfyUI-Nodes, then restart.

Hand installs hitting ModuleNotFoundError: cv2? pip install -r custom_nodes/AUN-ComfyUI-Nodes/requirements.txt.

Common issues

Where people get burned: treating all five outputs as "the result" when they're stages - only Both upscaled is the full chain, so save that one unless you tuned otherwise. Second, upscaling in latent space is not free of artifacts; if you see tiling or repetition, the fix is usually a smaller ratio or a lower upscaling_denoise, not more steps. And if you want to run the second pass on a different model, that's what the v4 sibling (AUNKSamplerPlusv4) is for - this node uses one model for both.