h4 - Complete Loader

H4_CompleteLoader is h4_Live's "declutter the left side of your graph" node. A stock ComfyUI workflow starts with a wall of loaders: checkpoint, VAE, CLIP, LoRA, and four separate Load Image nodes if you're doing any kind of reference-image work. CompleteLoader collapses all of that into one node - it loads the model stack and up to four images and their masks, and hands you the whole bundle on a single set of outputs. If your canvas looks like a loading dock, this is the fork lift.

It's built on the pack's UniversalLoader core (that's the legacy one, category _Legacy; CompleteLoader is the current all-in-one), so the loading logic underneath is the same code that's been running the pack's loader family. What it adds is the image side: most workflows that load four reference images spend most of their wiring just moving those tensors around, and this node makes that part of the same "load everything" step.

The settings that matter

• load_mode - Checkpoint (Standard) is the everyday pick: one .safetensors checkpoint gives you model, CLIP, and VAE together. Diffusers (Component) is for when you're assembling a model from separate parts (unet, clip, vae) instead of a single file.
• ckpt_name, unet_name, vae_name, clip_name - depending on mode, you pick either the checkpoint or the components. VAE and CLIP default to Baked / None, meaning "use what's inside the checkpoint" - which is the right choice for most people most of the time.
• lora_name and lora_strength - an optional LoRA applied at load time, strength 1.0 by default. Handy, but if you're stacking several LoRAs, a dedicated LoRA loader gives you more control.
• image_1 through image_4 - reference images loaded straight into the node. Note the tooltip caveat: GGUF support for unet/clip only appears if ComfyUI-GGUF is installed, so don't expect .gguf filenames in those dropdowns on a fresh install.

The outputs

Eleven of them, which sounds like a lot until you realize they replace a dozen separate loaders: MODEL, CLIP, VAE, then four image/mask pairs - IMAGE_1/MASK_1 through IMAGE_4/MASK_4. The masks are the quiet hero: if your reference image has a companion mask (say, a subject cutout), it comes out the corresponding MASK output ready to feed inpainting or detail work. If you leave an image slot empty, its outputs are None and downstream nodes that don't need it just ignore it.

Installing

ComfyUI Manager → search "h4_Live", or:

cd ComfyUI/custom_nodes
git clone https://github.com/m3rr/h4_Live

then restart. Core loader nodes are plain Python with no heavy deps. One pack-level caveat worth knowing: h4_Live is a one-person project still in active development, and the loader family is where the pack's "we refactored the naming" churn shows - there's a _Legacy UniversalLoader sitting right next to this one. If a workflow you downloaded references the old name, that's why.

Where people get burned

Don't wire an image slot expecting a mask to magically exist - MASK_n is only meaningful if the loaded image has one. And remember the tooltip on image_upload-style nodes elsewhere in the pack: if an image-fed grid comes out as colorful noise, the image often didn't load. Same rule applies here: empty slots are fine, but a slot you think is loaded and isn't will fail silently downstream.