Checkpoint / VAE Loader

The two-thirds version of this pack's Checkpoint / VAE / LoRA Loader - same idea, minus the LoRA slot, for the very common case where you're just running a bare checkpoint and don't need a LoRA in the mix. Stock ComfyUI splits this into two nodes (Load Checkpoint, then optionally Load VAE if you want to override the baked-in one); this collapses both into one and, same as its LoRA-equipped sibling, hands back the model's and VAE's filenames and hashes as strings alongside the actual model objects.

Mechanism. Loads the checkpoint file, splits it into MODEL and CLIP, and swaps in whichever VAE you've selected in place of whatever's baked into the checkpoint. Nothing exotic here - it's the same load path ComfyUI's own checkpoint loader uses, just paired with a VAE loader and a name/hash lookup in one box instead of two or three separate nodes.

Inputs and outputs. Two required dropdowns, ckpt_name and vae_name, populated from your models/checkpoints and models/vae folders. Seven outputs: MODEL, CLIP, and VAE for the actual generation pipeline, plus model_name, model_hash, vae_name, and vae_hash as strings - wire those last four into Save Image (Advanced)'s matching optional inputs if you want your saved files to carry real provenance data instead of "unknown."

Installing it. Comes 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. It reads whatever's already in your model folders - nothing to download for the node itself.

Troubleshooting.

• Empty ckpt_name/vae_name dropdowns mean ComfyUI can't see your models folder. Check extra_model_paths.yaml if your models live outside the default tree, and restart - not just refresh - after adding new files, since the combo lists build at startup rather than live.
• Grey, washed-out, or oddly desaturated output after picking a VAE almost always means a mismatch: you've attached a VAE the checkpoint wasn't trained against. Most current SDXL fine-tunes bake their VAE in and want this left on the checkpoint's own choice rather than forced to an external file - only reach for a separate VAE if the checkpoint's own page tells you to, or if you're on an older SD1.5 checkpoint that explicitly expects the standalone 840k VAE.
• Black output or NaN errors during decode on an SDXL checkpoint in fp16 is a known issue with the original SDXL VAE, not something this node introduces - the community fix is a rescaled fp16-safe VAE file (commonly distributed as sdxl-vae-fp16-fix), not forcing fp32 decode via a launch flag.
• If you only need this node for its metadata strings and are already happy with a stock checkpoint + VAE loader combo, there's no strict need to switch - the value here is consolidation and provenance data, not new generation capability.