AUN KSampler 2-Model

AUN KSampler 2-Model (AUNKSamplerPlusv4) is the same progressive two-pass sampler as KSampler Plus - base pass, latent upscale, second pass, optional pixel upscale, optional refine - with one addition that changes how you think about it: an optional second model that powers the latent-upscale pass. The idea is real and popular in hi-res-fix circles: generate with a fast or compositional model, then let a detail-oriented model do the second pass. If you've ever hand-built that with two KSamplers and a latent upscale in between, this is the one-node version.

How it works

The mechanics are the v3 chain: first pass → latent upscale (ratio, bicubic recommended) → second pass → Both upscaled via decoded pixel upscale → optional Refined image. The new input is model_refine, an optional MODEL connection. Wire a second model in and the latent-upscale pass uses it instead of the base model; leave it unconnected and the base model handles both passes, so the node degrades gracefully to a plain KSampler Plus. That "optional, falls back safely" behavior is why it's worth using even when you don't yet have a second model - you can wire one in later without restructuring the graph.

All the v3 controls carry over unchanged: steps_total / steps_first, start_step_second, cfg and cfg_latent_upscale (separate CFG per pass), latent_upscale / ratio / upscaling_denoise, the pixel-space image_upscale / image_upscale_method / image_upscale_model / image_upscale_ratio, and image_upscale_refine + img_refine_steps + img_refine_denoise for the optional final pass. Outputs are the same five images plus LATENT and the Upscaled type string for filename honesty.

When the second model actually helps

The community pattern: the base model composes (fast, low-CFG), the refine model adds detail (a detail LoRA'd checkpoint, a realism-tuned model). The tip is to keep the second pass denoise moderate - 0.5–0.7 per the tooltips - and remember that above ~0.6 the second pass starts changing composition, not just refining it. Pairing a distilled/turbo base with a full-quality refiner is the most common successful combo.

Installing it

Same pack, one install:

• ComfyUI Manager: search "AUN ComfyUI Nodes", install, restart.
• Manual: cd custom_nodes && git clone https://github.com/loz2754/AUN-ComfyUI-Nodes, then restart. Manual installs and ModuleNotFoundError: cv2? pip install -r custom_nodes/AUN-ComfyUI-Nodes/requirements.txt.

Common issues

Two-model setups multiply the usual tuning surfaces, so the failure mode is usually a mismatch: the second model has its own CFG habits (distilled models want CFG near 1–2, full models want more), and cfg_latent_upscale defaults won't know which you're using. Set it for the refine model, not the base. And if the second pass drifts off-composition, lower upscaling_denoise before changing models - you're more likely fighting denoise than the wrong refiner.