Load DiffDIS model by name

Every other model in this pack is a segmentation network that spits out a probability map. DiffDIS is not that. It's a fast diffusion model - SD Turbo, the 1–4 step one - repurposed for dichotomous image segmentation, and it processes two latents at once: one for the foreground mask and one for the edges. That makes it the only model in ComfyUI-RemoveBackground_SET that outputs a real edge map, and it makes it the slowest, heaviest thing in the pack by a mile.

The numbers tell the story. On the author's RTX 3060, DiffDIS Base F16 takes about 3.1 seconds and 6.1GB of VRAM at 1024px - compare BiRefNet's 516ms and 1.6GB. It also breaks the usual fp16 rule of thumb: the fp32 version is actually lighter on VRAM (5.6GB) than the fp16 one, which nobody fully explains. This is a "try it because it's interesting" node, not a "standardize your product pipeline on it" node.

The part everyone trips on: it needs a VAE

Unlike the other auto-downloaders, this one has two extra inputs, because DiffDIS is a diffusion model:

• vae (VAE) - required in practice. The model expects the SD Turbo VAE, so hook up a Load VAE node pointing at it. The README flags that it's the one download this pack might not do automatically.
• positive (CONDITIONING) - marked experimental. The model was trained with empty conditioning text; leave it unconnected.

You also get the standard model_name dropdown (just two entries: DiffDIS Base F16 at 1.8 GiB and DiffDIS Base at 3.5 GiB) plus device and dtype.

What comes out

• model (SET_REMBG) - feeds the same processing nodes as every other loader, and now they'll return edges masks, which only DiffDIS produces.
• train_w, train_h - both 1024×1024. DiffDIS is more restrictive about input sizes than the rest of the pack, so respect these.
• norm_params - for manual pre-processing.

Install

Same pack install as the other loaders: ComfyUI Manager search "Remove Background (SET)", or clone into custom_nodes and pip install -r requirements.txt. ComfyUI 0.3.48+, and if seconohe errors, pip install git+https://github.com/set-soft/seconohe.git for a fresh copy. Weights land in ComfyUI/models/rembg.

Who is this for? Edge-mask workflows - if you want the outline of the subject as a mask (for line art, edge-guided compositing, or just to see what a diffusion model thinks an edge is), DiffDIS gives you something nothing else here can. If you just want clean cutouts, pick BiRefNet and save yourself three seconds a frame. The honest framing: it's a party trick with real uses, and this loader is the lowest-friction way to try it.