DPR KSampler Advanced

What it actually does

Here's the thing that trips everyone up about Gaussian Shading watermarks: the watermark isn't in the image. It's in the noise. This node is a drop-in clone of ComfyUI's built-in KSamplerAdvanced with one extra input, watermarked_latent_noise, and one extra toggle, use_dprw_noise. Flip the toggle to enable and it starts sampling from your noise tensor instead of ComfyUI's random seed noise. That's the whole trick.

It looks identical to a normal generation because the "watermarked noise" is statistically indistinguishable from real Gaussian noise - the core promise of the Gaussian Shading paper (CVPR 2024) that this pack reimplements. Your output is supposed to be on par with a normal run, which is what makes it interesting versus older pixel-space watermarks.

How it works

Under the hood it's a thin wrapper over ComfyUI's own comfy.sample.sample. With use_dprw_noise on and a watermarked latent connected, it swaps the noise ComfyUI would normally synthesize from a seed for the tensor coming out of DPR_GS_Latent. Everything else is stock KSamplerAdvanced semantics: start_at_step / end_at_step for partial denoise, add_noise, return_with_leftover_noise.

The inputs that actually matter:

• use_dprw_noise - the master switch. enable means "use my noise, not yours."
• watermarked_latent_noise - wire the LATENT output of DPR_GS_Latent in here. This is the noise your watermark is baked into.
• sampler_name, scheduler, steps - keep these deterministic. euler, dpmpp_2m and DDIM are all fine. Ancestral samplers (euler_ancestral, dpmpp_sde) inject fresh randomness mid-trajectory, which scrambles the watermark trace - extraction has to walk the exact same trajectory backwards.
• add_noise - the example workflows leave this at disable. With use_dprw_noise on, the watermarked noise already plays the role of the initial noise, so this toggle becomes mostly moot.

Output is a single LATENT that feeds straight into VAEDecode.

Where people get burned

• noise_seed is decorative while use_dprw_noise is on. The seed that actually controls the noise lives in DPR_GS_Latent. If your image changes between runs, that's the seed you're looking for, not this one.
• Write down the sampler, scheduler and step count. Extraction via DPR_Extractor only works if the inversion uses the same settings - get them wrong and the recovered "watermark" is garbage.
• A shape mismatch between watermarked_latent_noise and latent_image produces a confusing torch dimension error mid-graph. Keep DPR_GS_Latent's width, height and channels in lockstep with your EmptyLatentImage and model.

Install

This is one of three nodes in ComfyUI-GaussianShadingWatermark, an unofficial reimplementation of the Gaussian Shading paper by GitHub user lthero-big. One install gets you all three (this sampler, DPR_GS_Latent and DPR_Extractor):

cd ComfyUI/custom_nodes/
git clone https://github.com/lthero-big/ComfyUI-GaussianShadingWatermark.git

Restart ComfyUI after cloning, or use ComfyUI Manager and search "GaussianShadingWatermark". Dependencies are a handful of common scientific packages (scipy, cryptography, numpy, Pillow) that ComfyUI mostly ships with already - no model downloads, no API keys. It's a hobby-scale research pack, so treat it as "does the one trick, mostly works" rather than production-grade provenance tooling.