Inputs Diffusers Refine Basic

Some diffusion setups want a second pass with a different model - a main UNet to compose, a refinement model to add polish. AUNInputsDiffusersRefineBasic takes the split-model basic contract (standalone UNet + separate CLIP + VAE) and adds an optional second diffusion model for exactly that, while deliberately skipping the older save-prep outputs that clutter the full Inputs nodes. It's the node you reach for when your pipeline is "generate with A, refine with B" and you'd rather not run two separate loader + sampler stacks.

How it works

The loading is the same diffusion-style flow as AUNInputsDiffusersBasic: diffusion_name for the primary UNet, clip_name + clip_type for the shared CLIP, vae_name for the shared VAE. The new bit is refine_diffusion_name - an optional refinement UNet. Select 'None' and the main model is reused for refinement, so the node gracefully downgrades to a plain basic setup when you don't need two models. When you do pick a refine model, you get a second MODEL REFINE output (plus model name refine) to feed a second sampler pass.

The SpeedLoRA handling here is worth reading because it's specific: speed_lora_full_both applies the full strength to both the main and refine models, while speed_lora_ratio splits the strength - the main model gets that fraction, the refine model gets the remainder. So you can give the compositor most of the SpeedLoRA boost and the refiner the rest, or everything to both. If that sounds fiddly, that's because it is - but it's also the difference between "refiner fights the LoRA" and "refiner cooperates with it."

Inputs and outputs

Inputs: diffusion_name, refine_diffusion_name, clip_name/clip_type, vae_name, the SpeedLoRA group, sampler/scheduler/cfg/steps, aspect_ratio/width/height/aspect_mode, batch_size, seed, megapixels, multiple.

Outputs: MODEL, CLIP, VAE, MODEL REFINE, model name, model name refine, sampler, scheduler, cfg, steps, latent, width, height, seed, batch size, speed lora ratio (the actual ratio applied, handy to feed a filename or display node).

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

clip_type mismatches are the usual first failure - the CLIP architecture has to match the CLIP file. On the SpeedLoRA side, if your refinement pass looks overcooked or the two models fight each other, that's usually the split between speed_lora_full_both and speed_lora_ratio - try dropping the refiner's share before changing anything else. And as with every Basic node here, there are no save-prep outputs; plan your filename building on the save side.