TryOff Anyone Node

Some ComfyUI nodes are general-purpose tools; this one does exactly one thing. TryOffAnyoneNode is a wrapper around ixarchakos's TryOffAnyone, a Stable Diffusion 1.5 inpainting fine-tune whose whole job is reverse virtual try-on: you mask the clothes on a person, and it regenerates that region as if the garment weren't there. The name tells you everything. There's no prompt, no negative prompt, no text box at all - the mask is the instruction.

That's genuinely rare, and it's the reason this node earns a spot in a workflow. Try-on (dressing a person) got plenty of attention - CatVTON, IDM-VTON, and Flux Fill-based pipelines all made the rounds. The reverse direction, undressing, mostly lives in uncensored inpainting territory, and this is one of the few purpose-trained open models for it. Because it's SD 1.5 under the hood, it's also light: per the KB, SD 1.5 still runs happily on 6 GB cards, and this fine-tune is no exception.

How it works

This is the part that surprises people. The node loads the stock SD 1.5 inpainting UNet and the sd-vae-ft-mse VAE from HuggingFace, drops in the fine-tuned attention weights from ixarchakos/tryOffAnyone, and then replaces every cross-attention layer with a no-op. Text conditioning is physically disabled - there is no encoder text, ever, no matter what you wish into a prompt box (there isn't one). The UNet just denoises the masked region guided by the surrounding visible pixels, using the standard 9-channel inpainting latent layout.

The guidance_scale dial is a side effect of that: with no text, classifier-free guidance runs between "predict with no visible context at all" and "predict using the visible image," so raising it makes the result cling harder to what's left on screen. Default 2.5 is a fine starting point.

The inputs that matter

Only two do the real work:

• image - your photo of a person.
• mask - where to regenerate. White (> 0.5) is the region that gets replaced, which is the standard ComfyUI convention, so masks from segmentation nodes work directly. The author's example workflow feeds it from LayerMask: SegformerB2ClothesUltra, which auto-segments clothing - that's the ergonomic path and worth copying.

Everything else is flavor. inference_steps defaults to 50, seed to 40. image_width and image_height are the actual output size - the node rescales both image and mask to those dimensions (max 512, step 8) before running, so your output is always exactly that size regardless of the source. The single IMAGE output wires straight into a Preview or Save node.

Two traps worth knowing. First, denoise is a lie: it's passed straight through as DDIM eta (stochasticity), not the denoise-strength you're used to from every other inpaint node. Leave it at 1.0 and don't expect "lower = subtler." Second, the mask gets binarized at 0.5 and upscaled with nearest-neighbor - any feather you painstakingly painted will be thrown away, so don't bother.

Installing

ComfyUI Manager finds it as TryOff Anyone; otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/RodrigoSKohl/comfyui-tryoff-anyone

Manager installs the requirements (diffusers, accelerate, transformers - none of which ComfyUI ships with). The real cost is the first run: the pipeline auto-downloads the SD 1.5 inpainting UNet, the VAE, and the fine-tuned weights from HuggingFace, several GB in total, into the HF cache (~/.cache/huggingface), not ComfyUI/models - that's where people go looking and don't find it. Also, the node builds its whole pipeline in __init__, so the download-and-load kicks off the moment the node enters the graph. A long first freeze is normal; subsequent runs are fine.

Bottom line

Reach for it when a mask is all you have - no prompt engineering, no LoRA hunting, a specific purpose-trained model that fits in low VRAM. It's a one-trick pony, but the trick isn't well served anywhere else. Give it a clean full-body shot and a decent clothes mask, and you'll be impressed at how far this 2024-era 1.5 fine-tune goes.