🛡️ 版权筛选器

If you generate images for clients or for a content pipeline, there's a recurring nightmare: a generation comes out beautifully - and it contains a character you have no right to ship. The classic case is trademarked IP: a model that knows "Totoro" or "Mickey" so well it happily renders them whenever the prompt drifts anywhere near. Catching that by eyeballing every image in a batch is exactly the kind of job that should be automated.

🛡️ 版权筛选器 (the display name is literally Chinese for "copyright filter") is an automated gate for it. It's an image → image node: you put it between your VAE decode and your save node, it screens the frame with a local vision-language model, and it has two outputs - clean, or blocked. The interesting part is what happens on "blocked": the image is not saved, and the frontend automatically re-queues the prompt with a new seed until the frame comes back clean (or it hits a retry cap).

How it works

The node uses a local Qwen3-VL (via the separate ComfyUI-QwenVL node) to look at the image and check it against your 版权黑名单 - a plain-text list, one IP per line, defaulting to Totoro, Mickey, Pikachu, SpongeBob, and Hello Kitty. There's also 也拦截其他知名IP (also block other well-known IPs), a boolean that widens the net beyond your explicit list. 检测精度 offers the same 4-bit / 8-bit / None (FP16) quantization choices as the pack's Ideogram4 node - 4-bit is the VRAM-friendly default.

When an IP is found, the node returns a blocked output via ComfyUI's ExecutionBlocker mechanism - which means downstream save nodes get nothing and the bad image never touches disk. At the same time it pushes a UI signal that the pack's JavaScript listens for. The JS then calls app.queuePrompt() to re-run the workflow with a new seed. The 最大重试 input (0–50, default 5) caps how many times it'll re-roll before giving up and alerting you instead of looping forever.

The one wire that makes it work

The retry loop only changes the image if your seed is actually randomizing. The node's own JS file says it in plain terms: set your seed node's control_after_generate to randomize, otherwise every re-queue regenerates the same image and the loop spins uselessly. That's the single most common setup mistake with this node, and it's entirely preventable.

Installing it

This one has a real dependency chain:

cd ComfyUI/custom_nodes
git clone https://github.com/LouisLU1997/ComfyUI-louis-use Louis_use

Then separately install ComfyUI-QwenVL (its own repo) and grab a QwenVL checkpoint. Without QwenVL, the node loads but the screening itself won't run - the pack's README says it "degrades gracefully," which means it passes things through, not that it works without the model. Start at 4-bit quantization to keep VRAM headroom.

The honest take

This is a niche tool with real requirements (a VLM, a model download, and a retry loop that needs a specific seed setting), and it's not a legal guarantee - a 4B vision model making a binary IP call will both miss things and cry wolf on lookalikes. But if you're running unattended batches where shipping a trademarked character is a real cost, an automated gate that re-rolls and skips the bad frame beats a human staring at 500 images. Set your expectations accordingly and it's genuinely clever. It's also an unusual node in that its frontend behavior (auto-requeue) is a feature you either want badly or should turn off - that's what 最大重试 0 is for.