TK 批量 LoRA 加载器

ComfyUI's stock LoRA loader gives you one LoRA. Two of them means a chain. Six means a noodle sculpture where changing a weight is a right-click-and-hunt expedition, and where a disabled LoRA has to be bypassed rather than switched off.

TK 批量 LoRA 加载器 is the pack's flagship node for exactly that mess: paste in a batch of <lora:name:weight> tags, get an editable list with per-LoRA switches, get trigger words out the other side. It's the node that TK Toolkit (anima-toolkit by 时运tk) is built around - the Civitai-integrated local LoRA manager is its companion panel.

Inputs and outputs

The required inputs are minimal, which is the point:

• model - your loaded MODEL. Nothing downstream changes; this node sits where a LoRA loader would.
• lora_syntax - the text field you live in. Paste <lora:char_x:0.8> <lora/style_y:0.6> and it parses into the visual list; edit the list and the text updates back. Two-way, so you can work from either side.
• clip - optional. Connect it if you also want the LoRAs applied to the text encoder.
• output_trigger_words - a master switch, on by default. Turn it off and the third output returns an empty string. LoRA loading is unaffected.

Outputs are MODEL, CLIP and trigger_words. The first two go where your chain expects them. The third is trigger words for the LoRAs that are currently active - weight non-zero - with one deliberate quirk: a LoRA whose file is missing but which is switched on still contributes its words. That's helpful when you're building a prompt before you've downloaded everything.

What the node panel gives you

This is where it earns the name. Attach your LoRA folders and the node's browser walks them recursively, keeping relative paths (anima base/style.safetensors), so two files named the same in different folders don't collide. If the toolkit has matched your files against Civitai by SHA256, you see the preview images right there and click a thumbnail to add it to the node.

Per LoRA you get a switch (off means it stays in the list but stops contributing - no bypassing), a weight from −2 to 2 with scrubbing on the angle brackets (drag horizontally for continuous change, click for 0.05 steps), drag-to-reorder, and delete. Negative weights are allowed and occasionally useful. You can also save a 组 - a named preset of the current set - and switch between groups in one click.

The 验证标签 button checks whether the LoRAs in your text actually exist locally and offers to copy the names of the missing ones so you can search Civitai for them. The 更新 button checks the pack for updates, including commit-level changes within the same version number.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/Ararararararaki/comfyui-anima-toolkit
# restart ComfyUI

Or ComfyUI Manager, search TK Toolkit (registry id anima-toolkit). Only aiohttp and requests are required - the panel is shipped prebuilt, so no Node.js, no npm install. After updating, restart ComfyUI; if only the frontend changed, a hard browser refresh (Ctrl + Shift + R) is enough.

Civitai features need network access, and the local manager stores its scan data in the plugin's data/ directory plus the browser's IndexedDB. That last part has teeth: open the panel and ComfyUI from the same host and port. localhost:8188 and 127.0.0.1:8188 are different browser origins, and IndexedDB content - categories, custom thumbnails, prompt library - will look like it vanished if you switch between them.

Gotchas

Missing LoRA doesn't always fail loudly. The loader resolves names against your model paths. If a tag doesn't match, check the filename first: fuzzy matching exists for cases like a real file named thing (1).safetensors, and if it warns about a 500 from /anima/lora/info, update the pack - that path was broken and fixed in 2.9.0.

Applying LoRAs to CLIP is optional for a reason. Anima LoRAs are trained against a DiT with a Qwen3 text encoder, and the interesting part of the model is the diffusion side. Wire clip when a LoRA was trained with text-encoder keys; leave it dangling otherwise.

Trigger words are a convenience, not a requirement. Anima reads natural language and character names well, and the community consensus on it is that a lot of what SDXL users trained LoRAs for, it already knows. Check animadex.net for whether the model knows your character before you assume you need a LoRA and its triggers at all.