Boyo VAE Decode

Here's the thing nobody tells you about BoyoVAEDecode: it's not just a VAE decode. It's a VAE decode with a built-in content filter. Every frame gets run through an NSFW image classifier, and anything flagged as NSFW above a 0.9 confidence score is silently replaced with a placeholder image - a static example PNG shipped with the pack. You get your image back, it just isn't the image you generated.

The README lists this node as a plain "Boyo VAE Decode" with no mention of the filter. That's a gap worth knowing about before you spend an afternoon convinced your sampler is broken.

How it works

The node's constructor loads a HuggingFace image-classification pipeline - Falconsai/nsfw_image_detection - the first time it runs (so it needs an internet connection and a working transformers install on that first use). When you feed it samples (LATENT) and vae, it decodes normally, then loops over every frame, runs the classifier, and for any frame scoring over 0.9 it overwrites that frame with images/example.png resized to match. Output is the IMAGE, censored or not.

The threshold is hardcoded at 0.9 - there's no setting to tune it, and no bypass flag.

Where it fits

Honestly: carefully. If you're generating anything the classifier might call NSFW, this node will eat your output without telling you. If you're doing clean content and want a single decode node that also keeps the graph free of accidental nudity (say, for a public-facing or shared pipeline), the filter is doing you a favour. But it's a side effect, not a documented feature - so the safer default for most people is to use the standard VAEDecode node from ComfyUI core, which has zero surprises.

There's also a Boyo Tiled VAE Decode in the pack (not this node) for large-image tile decoding - don't reach for this one expecting tiling; it doesn't tile.

Installing it

Ships with Boyonodes (DragonDiffusionbyBoyo):

cd ComfyUI/custom_nodes
git clone https://github.com/DragonDiffusionbyBoyo/Boyonodes

Restart ComfyUI or install "Boyonodes" via ComfyUI Manager. This node specifically needs transformers (already in the pack's requirements.txt) and pulls the Falconsai/nsfw_image_detection model from HuggingFace on first use.

Gotchas

• Silent replacement. The swap happens frame-by-frame with no obvious UI signal; if an output looks wrong, check the console logs and this node first.
• First-run download. No internet / blocked HuggingFace, and the node fails at init with a model-download error.
• No controls. Threshold, placeholder image, and the on/off switch are all fixed in code.

Useful as a guardrail in a family-friendly pipeline, but read the display name carefully - there's a reason this one needed an article. If your workflow suddenly started producing a single repeated image, this is the first node to suspect.