DeepFakeDefender_Sampler

This is the half of the pack that actually earns its keep. DeepFakeDefender_Loader sits around holding a model; this node scores your images, prints a verdict per image, and splits the batch into "looks fake" and "looks real" piles automatically. If you've ever needed to sort a folder of downloaded faces into suspect and clean before you let them anywhere near a training run, this is the tool.

The model behind it is the 1st-place solution to the Global Multimedia Deepfake Detection competition (Image Track) - a seven-expert ensemble of ConvNeXt-Tiny and EfficientNet backbones running on EMA weights. This sampler is just the inference front end: it resizes, runs the ensemble, averages the experts' probabilities, and compares the result against your threshold.

What it does step by step

For each image in the batch, the node:

1. Upscales or downscales it to your crop_width × crop_height (nearest-exact, center-cropped).
2. Applies the preprocessing pipeline that came out of the Loader's transform_val socket - ImageNet normalization, resized to the model's native 512×512.
3. Runs the ensemble and gets a single number: the probability the image is a deepfake, 0–1.
4. Prints that prediction to the console (in English and Chinese, if you care) and routes the image to one of two output piles.

The inputs that matter

• image - your input, an IMAGE tensor. Batches are handled, so you can feed multiple faces at once.
• net and transform_val - both come from DeepFakeDefender_Loader. Yes, transform_val is typed as MODEL; it's actually a torchvision transform pipeline. Wire it anyway.
• threshold - the decision boundary, default 0.5. Above it → classified deepfake, below → real. Raise it if you want fewer false alarms (more likely to call a fake "real"), lower it if you'd rather flag everything suspicious. This is the one knob you'll actually tune.
• crop_width / crop_height - default 512×512, which matches the model's native input. The README says cropping the image first gives a "slight improvement in accuracy." Leave them at 512 unless you have a reason.

The three outputs

• string - the text verdict, one line per image with its deepfake probability.
• above - the IMAGE output of images above threshold (i.e., flagged as deepfake).
• below - the images below threshold (judged real).

That's the genuinely nice part: wire above into one preview and below into another and you get visual sorting for free, no Python.

The trap

If a bucket comes up empty - say every image is real, so nothing lands in above - the node does not give you an empty tensor. It returns a 512×512 white image with the text "No image's prediction is above the X" drawn on it. Sounds harmless, until a downstream node chokes on a grayscale-looking image you didn't expect. If you're looping the output, check for that placeholder or keep it in mind when you see a white frame appear in your preview.

Also note the split uses <= for below and > for above, so a prediction exactly at threshold counts as "real."

Install & gotchas

Same pack as the Loader, so the install is identical:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_DeepFakeDefenders.git

Model weights (weight.pth + ema.state) must be downloaded from the README's Google Drive / Baidu links into ComfyUI/models/DeepFakeDefender/ - nothing auto-downloads. Two things worth knowing before you build a workflow on this:

• It's CUDA-only. The loader forces .cuda(), so CPU-only and Apple Silicon installs will fail at the Loader, not here.
• timm is needed but isn't in the requirements file (which is entirely commented out). If imports fail, pip install timm.
• The whole thing is CC BY-NC 4.0 - fine to tinker with, not for commercial use without checking the license.

Is it a silver bullet for deepfake detection? No - the author's own example images include false calls, and no threshold makes a classifier perfect. But as a free, local, batch-sorting first pass, it's surprisingly practical.