ModelSamplingWaifuDiffusionV

This node exists for exactly one job: making the WaifuDiffusion V test checkpoint (wdv-test-v0.1.safetensors) generate anything at all in ComfyUI. Without it you'll feed the model to a KSampler, wait, and get static or a mess of noise - because WDV doesn't sample like any SDXL model you've used before. The author's own reddit note says it plainly: different sampling than regular SDXL, the custom node is required, specific samplers are mandatory, and it will not work in A1111.

Think of it as a cousin to ComfyUI's built-in ModelSamplingEDM and ModelSamplingSD3 nodes. Those exist because some checkpoints were trained with a different noise schedule than the sampler's default assumes. ModelSamplingWaifuDiffusionV does the same trick for the WDV test model, which was trained on a hybrid of the SDXL beta schedule and the cosXL "tan" schedule from the rectified-flow research line. The community reaction when it dropped in mid-2024 was "incredibly rough currently, but glad to see teams actually training stuff" - this is a research peek, not a daily driver. Which is exactly why it ships with its own bespoke sampling code.

How it works

The node clones your model and swaps out its model_sampling with a custom object that blends two things:

• The noise schedule (the shape of the noise curve), which can be Tan (the cosXL-style tan-scaled curve), XLTC (SDXL's sigma curve up to a threshold, then a tan tail - this is the hybrid the model was actually trained on), or XL10 (plain SDXL-style log-linear sigmas, no tan).
• The prediction objective, either edm (Karras-style epsilon prediction, what waifu-diffusion 1.5 used) or v_prediction (velocity prediction, the SDXL-family objective that fixes the zero-terminal-SNR brightness problem).

The default is edm sampling + XLTC schedule, and that default is there for a reason: it's the combination the WDV team trained on. It works because the gist it's built from is 100% the work of neggles - the same person whose schedule code ComfyUI uses elsewhere - wrapped in a proper folder-based custom node.

The inputs that matter

Most of them, you leave alone. The three that actually matter:

• model - your loaded WDV test checkpoint, straight from a CheckpointLoaderSimple. Wire the output into a KSampler.
• schedule - keep XLTC unless you're experimenting.
• sampling - edm is the default and the safe pick.

sigma_min (0.001), sigma_max (1000), and tan_scaling (1.6, range 1–2) are advanced knobs that rescale the schedule. tan_scaling is the one you might actually touch - it controls how aggressively the tan tail kicks in - but for a test model, you're guessing blind. The output is a single patched MODEL, which then feeds your sampler.

Installing it

The pack itself is a one-file wrapper with zero extra dependencies - no requirements.txt, no models of its own. Install via ComfyUI Manager (search "WDV Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/BetaDoggo/ComfyUI-WDV-Nodes

Then restart ComfyUI. The real download is the model: grab wdv-test-v0.1.safetensors from the waifu-diffusion/wdv-tests repo and drop it in ComfyUI/models/checkpoints. The pack's example workflow loads exactly that filename, so rename if yours differs.

Where people get burned

• Patching the wrong model. This node is a one-trick pony for the WDV test checkpoint. Running your regular Pony or Illustrious checkpoint through it and getting garbage isn't a bug - the node patched in a schedule that checkpoint wasn't trained on.
• Ignoring the sampler. The example workflow uses dpmpp_2s_ancestral with the simple scheduler. Random samplers produce broken images, and neggles himself flagged that some samplers just don't play nice with wdV's cosXL-based schedule.
• Expecting production quality. It's a v0.1 test trained for roughly five epochs on ~4.8 million anime images. It's a fascinating look at the WaifuDiffusion team's next-gen direction - the "V" line - not a replacement for your current anime model. Judge it accordingly.

If you're here because you saw the model on a checkpoint list and want anime images today, skip the detour. If you're curious what the next waifu-diffusion could look like, this is the node that unlocks it.