加载分镜 LoRA · Load Storyboard LoRA

Here's the normal situation. Your storyboard's character cards know which LoRA each character uses, at what strength. The loader node surfaces that as lora_name and lora_strength strings. A native LoraLoaderModelOnly takes its lora name from a dropdown of files it scanned at startup, so it can't accept a string from upstream: you end up manually matching the file every time you switch pages. SodaLoadStoryboardLoRA is the missing adapter - model in, string name and float strength in, patched model out.

What it does

Take the model from your UNETLoader stack, pass it a name and a strength, and it patches the model weights and hands the model back. The mechanics are ordinary ComfyUI LoRA loading: it resolves the file, loads the torch file, and applies LoRA to a model only - no CLIP side, which is right here because MOSS's text encoder isn't a LoRA target.

Two behaviours are worth knowing, and both are in the source rather than buried in docs.

First, an empty lora_name returns the model unchanged. That's a deliberate pass-through so you can drop this node into a chain and leave it inert when a page has no character LoRA - the loader's suggestion comes back blank and nothing happens. The pack's showcase page is LoRA-free by design, character defined by text alone, so this path is the common one.

Second, the name is strictly a path relative to models/loras. Absolute paths and .. are rejected outright, and the character card's lora.path is expected to look like characters/hero.safetensors. Then it asks ComfyUI's folder registry for the full path and raises if the file isn't there. No silent skip, no falling back to the base model. That failure is loud on purpose: quietly generating a page with the wrong character is worse than an error.

Inputs and output

model is required. lora_name is a string, empty by default. strength is a float defaulting to 1.0 with a startling range of −100 to 100 - you'll live between 0.5 and 1.2 in practice, but negative values are occasionally how people subtract a style.

One output: model. Wire it between your UNET loader and whatever sampling node you use, and wire lora_name and lora_strength from the loader. Done - the storyboard now decides which character LoRA is active per page instead of you remembering.

Whether you should is another question. Character LoRAs are the standard tool for keeping a face consistent across panels, and they're also the thing that bleeds into everything else in the image. The community's harder-won lesson is that no single consistency trick survives a whole book: the projects that finish one use a stack of LoRA, control, careful prompting and a lot of manual correction. This node gets the LoRA part out of your hands; it doesn't make the rest unnecessary.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/cOkieeman/ComfyUI-Soda-Storyboard
cd ComfyUI-Soda-Storyboard && python -m pip install --upgrade soda-storyboard

Manager → search Soda Storyboard (comfyui-soda-storyboard), or comfy node install comfyui-soda-storyboard, then restart. Character LoRAs go in models/loras, and if your cards reference a subfolder, keep the folder structure - the name resolves relative to that root.

Where people get burned

"找不到 LoRA 文件" is the expected failure when the card suggests something you don't have. Either drop the file in the right place under models/loras, or clear lora_name (or bypass the node) and let the prompt alone carry the character.

The pack warns once, not repeatedly, when a page has several characters that each carry a LoRA: one at a time is the intent, and it still lets the page compile. Loading two character LoRAs at once is how you get two faces fused into one, and no node is going to stop you.

If the suggested name has a Windows-style backslash or a leading /, fix the character card rather than the node - the path policy is checked before the file lookup, so a bad path fails even when the file exists.

And if you're using the default example workflows, remember they ship with the LoRA group bypassed and the LoRA loader unwired. You have to swap the native LoraLoaderModelOnly out for this node yourself; nothing auto-adopts the suggestion.