Inputs Refine

AUNInputsRefine is AUNInputs with a second model attached. It loads the main checkpoint, creates the latent, sets all the sampler parameters, and prepares the save filename - and then also loads an optional refinement checkpoint, giving you a second MODEL output for a refinement pass. For workflows that end with a polish pass using a different model (a detail-oriented refiner, a different flavor of the same base), this collapses what would otherwise be two loader stacks plus a pile of parameter nodes into one.

What it adds over AUNInputs

Two things. First, refine_ckpt - an optional second checkpoint, select 'None' to skip it (in which case the node behaves like plain AUNInputs with an extra, unused output). When set, you get MODEL REFINE and ckpt name refine outputs to drive a second sampler pass. clip_skip applies to both models.

Second, SpeedLoRA gets smarter. Beyond the usual speed_lora / speed_lora_model / speed_lora_strength, there's speed_lora_full_both - apply the full LoRA strength to both main and refine models - and speed_lora_ratio, which splits the strength between them: the main model gets that fraction, the refine model gets the remainder. That split is the feature, because a SpeedLoRA tuned for the compositor can easily fight a refiner. Give the refiner less of it, or hand it everything, per workflow. The node even outputs the resolved speed lora ratio so you can see (and log) what actually got applied.

Inputs and outputs

Inputs: ckpt_name, refine_ckpt, the SpeedLoRA group, clip_skip, sampler/scheduler/cfg/steps, aspect_ratio/width/height/aspect_mode, batch_size, seed, megapixels/multiple, and the save-prep set (MainFolder, ManualName, name_mode, prefix, date_format, crop, words).

Outputs: MODEL, CLIP, VAE, MODEL REFINE, ckpt name, ckpt name refine, sampler, scheduler, cfg, steps, latent, width, height, seed, MainFolder, Filename, prefix, date format, batch size, name mode, speed lora ratio.

Installing it

One 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

The README flags a specific migration gotcha here: the SpeedLoRA-related input set changed, so if you load an older workflow that uses this node, "the SpeedLoRA-related widgets may need to be checked or reconnected." That's the first thing to check when an old graph loads with odd widget states. Beyond that, the usual two-model confusion applies - if your refinement pass looks wrong, the speed_lora_ratio split is the first lever to pull, before touching denoise on the refiner. And as a full Inputs-style node, it carries the legacy save-prep contract; the pack's forward-looking recommendation is Basic nodes plus AUN Save Image V2.