Josia模型加载

JosiaCheckpointPlus (Josia模型加载) is the flagship of the JosiaNodes pack and the one with the most ambitious claim in the README: it's a "100% replacement" for ComfyUI's CheckpointLoader, UNETLoader, CLIPLoader, and VAELoader, all in a single node that auto-detects what kind of model you feed it. You pick a file, it figures out whether it's an all-in-one checkpoint, a standalone UNET, or a GGUF quant, and wires up the right loading path - including, in the video-model era, a second VAE port for audio.

The auto-detect, in plain terms

Drop a model into main_model and it classifies it into one of three shapes:

• AIO checkpoint - the file bundles UNET + CLIP + VAE (the classic sd_xl/flux safetensors). It reuses the built-in components and grays out the external CLIP/VAE selectors, because they'd conflict.
• Standalone UNET - a bare UNET file. You freely pick the CLIP and VAE to pair with it.
• GGUF UNET - a quantized model. This is where the pack's only real external dependency lives: GGUF files need the ComfyUI-GGUF plugin (city96's, the standard one) installed, because decoding the quant is that plugin's job. The node itself handles the smart scheduling and hands off to the plugin's UnetLoaderGGUF/CLIPLoaderGGUF classes. Non-GGUF files (ckpt/safetensors/bin) run completely standalone.

The inputs that matter

• main_model - the file picker; auto-detection kicks off from the selection.
• clip_type - a 36-choice dropdown mirroring the official CLIPLoader types (sdxl, flux, flux2, sd3, wan, qwen_image, joyimage, mage, minimax, etc.). AIO mode sets it automatically; standalone-UNET mode needs you to pick the one matching your CLIP file, and unselected falls back to stable_diffusion.
• clip_name and vae_name - external CLIP/VAE selection, only active in standalone/GGUF modes.
• vae2_name - the second VAE, aimed at video models that carry both a video VAE and an audio VAE (LTX, MMAudio, SA3, MiniMax H3). It auto-converts official audio_vae./vocoder. weight prefixes so you don't have to rename files, and it lists both models/vae and models/checkpoints because audio VAEs often sit in the latter.
• lock_unet_vram (default on) - the "keep the UNET alive" switch. It preheats the UNET to the GPU on load and keeps it resident across reruns, so reusing a workflow skips the reload. The tooltip is honest that new ComfyUI already keeps models resident after first use - this mainly gets you the preheat and guaranteed residency, without force-holding VRAM.

Outputs

Four ports: MODEL, CLIP, VAE1, and VAE2. Same shapes ComfyUI's loaders always emitted, so it drops into an existing workflow as a straight swap - that's the entire compatibility pitch.

Install

The pack itself is dependency-free:

cd ComfyUI/custom_nodes
git clone https://github.com/Josia-doit/ComfyUI_JosiaNodes

or via ComfyUI Manager (search ComfyUI_JosiaNodes), then restart. If you plan to load GGUF models, you also need:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF

and restart again. The README's diagnostic: if the node says it can't find a GGUF loader, the plugin isn't installed or isn't enabled - the node only schedules, it doesn't decode.

Honest takes

This is the most ambitious node in a pack that's otherwise small utilities, and it's genuinely clever about the video-era stuff - dual VAE support with automatic prefix conversion is a real convenience nobody else's one-node loader bothers with. But the "100% replacement" framing is doing work. If you mostly load one standard checkpoint, the stock CheckpointLoader has one dropdown and no learning curve; this node's 36-type CLIP dropdown is power you don't need. Where it earns its keep: mixed workflows hopping between AIO checkpoints, standalone flux UNETs, and GGUF quants, where remembering which loader each file needs is exactly the kind of bookkeeping a single auto-detecting node should kill. Just be ready for the Chinese UI and the busy panel - and know that if a GGUF load errors, the plugin, not this node, is where the bug usually lives.