MinusZone - KolorsCheckpointLoaderSimple

Kolors ships in two packaging flavors, and MZ_KolorsCheckpointLoaderSimple is the loader for the second one. If you grabbed the official model from HuggingFace you have the diffusers unet/ layout, which wants MZ_KolorsUNETLoaderV2. But a lot of Kolors releases - especially merged/derivative checkpoints that people trade as single files - come as one .safetensors you're supposed to drop in models/checkpoints/. That's what this node loads. It's the pack's answer to "what if my Kolors file lives in the checkpoints folder."

How it works

The node calls ComfyUI's load_checkpoint_guess_config - the same guess-the-architecture loader that powers the core CheckpointLoaderSimple - but inside the pack's Kolors hook context, so the loaded model is recognized as the Kolors SDXL-style UNet with its encoder_hid_proj intact. Two outputs, MODEL and VAE. Notably, it loads no CLIP (output_clip=False): Kolors has no CLIP text encoder to load, because its text side is ChatGLM3. That's not a bug - it's the architecture. If your graph expects a CLIP output from this loader, that's where the confusion starts; the text path runs through MZ_ChatGLM3Loader + the V2 encode node instead.

Inputs and outputs

• ckpt_name - dropdown of everything in models/checkpoints/.
• Outputs: MODEL (→ KSampler) and VAE (→ VAE Decode).

Install

Standard pack install: ComfyUI Manager search "ComfyUI-Kolors-MZ", or git clone https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ into custom_nodes and restart. Then put your checkpoint-format Kolors file in models/checkpoints/. The pack's FAQ is refreshingly pragmatic about the whole two-flavors business: if you can't tell which type your model is, just try both loaders. This one and MZ_KolorsUNETLoaderV2 cover the two cases, and one of them will take your file.

Common issues

The failure modes are the mirror image of the UNet loader's: if this node can't detect the model type, your file is probably the diffusers unet/ layout, in which case use MZ_KolorsUNETLoaderV2 (the README suggests trying the UNet loader first for files from HuggingFace's unet folder). And if you're wondering where the CLIP output went - there isn't one, by design; wire text conditioning from the ChatGLM3 nodes. Also remember that this loader rides core ComfyUI's guess-config logic, so an outdated ComfyUI can throw model_detection errors; update core before debugging anything else.