Krea2 Merge • Load LoRA

The name is a lie, and it's a useful lie to understand. Krea2 Merge • Load LoRA loads a LoRA, yes - but what comes out of its output socket is not a model. It's a raw state dictionary wearing a "MODEL" badge so that ComfyUI's sockets stay tidy. Wire it into a KSampler and nothing good happens. Wire it into the pack's Merge LoRAs node and you have the entire input side of the merge pipeline in one step.

This node exists for exactly one job: get a LoRA file off disk and into the merger without loading or modifying the Krea 2 base model. That's the pack's whole philosophy, and it's what makes the merger fast and VRAM-cheap - nothing touches a 12B diffusion transformer.

How it works

It reads the file straight from ComfyUI/models/loras. .safetensors files go through the safetensors loader; anything else (torch.load formats) goes through PyTorch's own path. Both land in a CPU state dict, and that dict is what gets passed along - a dictionary of tensors keyed like lora_A/lora_B (PEFT/Diffusers, which is what Krea 2 LoRAs ship as) or lora_down/lora_up (Kohya). Because Krea 2 LoRAs use the PEFT key convention, the classic Kohya-era merge tools silently ignored them; this fork was built to read both.

The inputs that matter

Only two, and you'll touch mostly one:

• category_filter - a folder dropdown ("All" plus any subfolders under models/loras). The front-end filtering is handled by the pack's small JS file, so this is pure convenience.
• lora_name - the actual file picker, already filtered by the dropdown above it.

The single output, named model, is the state dict in disguise. Treat it as a black box and you'll be fine: it only plugs into another node in this pack.

Install

Identical to every node in the pack - install once, use all four. In ComfyUI Manager, search Krea2 Merge (it's on the Comfy Registry as comfyui-krea2-merge), or clone:

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/comfyui-krea2-merge

Restart ComfyUI after. No model files to download, no extra Python packages beyond safetensors, which is already present in a standard install. Because the fork uses its own Krea2Merge_ node IDs, it happily coexists with the original comfyui-merge extension - both can be installed and neither clobbers the other.

Gotchas

The big one is the type fiction: if you drag this into a workflow expecting a usable model, you'll get errors downstream. This is a feeder node, not an application node. If you actually want to apply a LoRA to a model, that's the pack's Apply LoRA node (or stock ComfyUI). Also worth knowing: the loader itself is permissive, but the merger downstream is not - feed it a DoRA or LoCon file and Merge LoRAs will reject it with a clear error rather than quietly producing garbage, so don't be surprised when validation happens one node later.