Checkpoint / VAE / LoRA Loader

This is the loader half of the same idea TW-CUI-Util's Save Image (Advanced) is the save half of. Stock ComfyUI makes you chain a Load Checkpoint into a Load LoRA into, optionally, a separate Load VAE - and none of those nodes tell you anything about what you just loaded beyond the model/clip/vae objects themselves. Want the checkpoint's filename or a hash for your saved metadata later? That's more nodes. This one does checkpoint plus VAE plus LoRA in a single box and hands back both the usable model objects and their names/hashes as plain strings, specifically so you can wire those strings straight into a metadata-aware save node.

Mechanism. Under the hood it's doing the same three loads ComfyUI always does: read the checkpoint file, split it into MODEL and CLIP, swap in a VAE if you've picked one, and apply the LoRA's weight deltas to both the model and the CLIP with independent strength values. What's different is that it does all three loads in one place and additionally names/hashes each component for you, instead of leaving that as a separate step you'd otherwise have to build yourself.

Inputs and outputs. Three required dropdowns - ckpt_name, vae_name, lora_name - populated from whatever's sitting in your models/checkpoints, models/vae, and models/loras folders. Two float sliders, lora_str_model and lora_str_clip, both defaulting to 1.0 with a -100 to 100 range (in practice you'll live in roughly 0-1.5; more on that below). Nine outputs in total: MODEL, CLIP, and VAE for wiring into your sampler and prompt nodes as normal, plus six strings - model_name, model_hash, vae_name, vae_hash, lora_name, lora_hash. Those six are the whole point of reaching for this node instead of the stock loaders: wire them into Save Image (Advanced)'s matching inputs and your output files carry real provenance data instead of the word "unknown."

Installing it. Ships with TW-CUI-Util. ComfyUI Manager: search "TW-CUI-Util", install, restart. Manual: git clone https://github.com/TW-CUI/TW-CUI-Util into ComfyUI/custom_nodes, restart. Nothing extra to download - it reads whatever checkpoints, VAEs, and LoRAs you already have in your standard model folders.

Troubleshooting.

• Empty dropdowns (no checkpoints/VAEs/LoRAs to pick from) almost always means ComfyUI is looking in the wrong folder. Check your extra_model_paths.yaml if you're pointing at models outside the default ComfyUI tree, and restart - not just refresh - after adding new files, since the combo lists build at startup rather than live.
• LoRA strength: the old SDXL-era rule of thumb is roughly 0.5-0.8, and both sliders here default to a flat 1.0. If a LoRA is overpowering your generation - one face showing up on everyone, a style bleeding into things that shouldn't have it - that's usually the first knob to turn down before you blame the checkpoint. lora_str_model and lora_str_clip are independent, so you can, say, keep model strength high for likeness while pulling clip strength down if the prompt's stopped listening to you.
• Grey or washed-out output after swapping in a VAE is the classic mismatched-VAE symptom. Most current SDXL fine-tunes bake their VAE in and don't want a separate one attached at all - if the checkpoint's own page doesn't explicitly recommend a VAE, leave this set to the checkpoint's baked-in choice rather than forcing one.
• The -100 to 100 range on the strength sliders is generous to the point of being a footgun. Anything much past roughly ±2 is going to blow the image out rather than make the effect "more" - there's no clamp here protecting you from typing in a wild number.