Draw Things LoRA

LoRAs are the small patch files that teach a checkpoint a style, a character, or a concept - tiny compared to the base model, but they change how the denoiser behaves. In the Draw Things pack, DrawThingsLoRA is how you stack them, and it's built to hold a lot: up to eight LoRA slots in a single node, each with its own weight and target.

The DT_LORA output feeds the sampler's lora input, and that's where the interesting part kicks in - the LoRA files themselves live in Draw Things, not ComfyUI. The dropdowns are populated from the server's model browser, and the sampler sends the file references over the wire so the server applies them during denoising. You never load a LoRA into ComfyUI's memory at all.

Standard disclaimer for this pack: it's deprecated. The README's first line says to switch to the official draw-things-comfyui extension. The nodes still work.

How it works

add_to_pipeline walks the node's eight slots, collects every LoRA you've selected, and builds a LoraStack - a list of {model, weight, mode} entries. The optional lora_stack input lets you chain nodes together, so you're not limited to eight; each node appends its stack to whatever came in. The sampler forwards the whole list to the server, which applies them to the diffusion and CLIP models.

Two settings per slot matter:

• weight - −5 to 5, default 1. Negative weights are allowed and legitimately used to subtract a style. Start at 1 and move in 0.1 steps.
• mode - All, Base, or Refiner. If you have a refiner wired up, this picks whether the LoRA applies to the base model, the refiner, or both. If you don't have a refiner, the tooltip says it plainly: this setting is ignored.

The pack ships an example workflow ("LoRAs, Upscaler and Refiner") showing exactly how LoRA nodes chain into the sampler alongside the upscaler and refiner, if you want a concrete template.

Installing it

ComfyUI Manager (search ComfyUI-DrawThings-gRPC) or:

cd ComfyUI/custom_nodes
git clone https://github.com/Jokimbe/ComfyUI-DrawThings-gRPC

Restart ComfyUI, and configure the server side - Draw Things app with API Server on, gRPC, TLS on, Model Browser on, Response Compression off (or gRPCServerCLI --no-response-compression --model-browser).

Troubleshooting

"LoRA not working" is the classic, and the causes are the same here as anywhere, with one twist. The twist: the file must be installed in Draw Things - the server only offers what its model browser can see, so a LoRA sitting in ComfyUI's models/loras folder simply won't appear in the dropdown. After that, the usual suspects apply: the LoRA needs its trigger word in the prompt (many style and character LoRAs do nothing without it), it was trained for a different base architecture than the model you picked (a Flux LoRA won't work on SDXL, and Pony/Illustrious/SDXL LoRAs often aren't cross-compatible despite sharing an architecture), or it's an anime LoRA that expects CLIP skip 2 while the sampler defaults to 1. Check the model page for the trigger word, then verify the wiring from lora_stack into the sampler - the stack silently stays empty if you never selected anything in a slot.