TSSAT

TSSAT ("Two-Stage Statistics-Aware Transformation", 2023) is the pack's slow-yet-quality middle ground: slower than CAST or EFDM, faster than Neural Neighbor, with a distinctive mechanism and an unusual knob. Like everything in ComfyUI-StyleTransferPlus, it's non-diffusion arbitrary style transfer - content and style images in, stylized image out.

The "two-stage" in the name is the whole story. Stage one is a patch-based statistics-aware transformation: it slices the content and style into local patches, then matches each content patch's statistics to the most similar style patches - which is how it keeps local structure while importing texture. Stage two runs the result through a trained decoder network that turns the transformed features back into a full image. The patch matching (with stride 1) is what makes it slow.

The knob nobody expects: max_iter

Here's the quirk that makes TSSAT stand out: the node runs the whole pipeline in a loop, feeding its own output back in as the new content. max_iter (default 1) controls how many times. This isn't an EFDM-style strength slider - bumping it up makes the stylization progressively stronger as the result is re-transformed again and again. Start at 1; if the style feels weak, try 2 or 3 and watch the texture accumulate. Each iteration costs real time, so there's a sweet spot somewhere around 2–3 before it just gets noisy.

Inputs and outputs

• src_img / style_img - content and style.
• size - square edge length after resizing (default 512).
• do_crop - resize-then-center-crop to square instead of squishing.
• max_iter - the self-reinforcement loop count, default 1.

Output: out_img (IMAGE) → Preview/Save.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/FuouM/ComfyUI-StyleTransferPlus

or ComfyUI Manager → search ComfyUI-StyleTransferPlus → install → restart.

Models - two files:

ComfyUI/custom_nodes/ComfyUI-StyleTransferPlus/models/tssat/decoder_iter_160000.pth
ComfyUI/custom_nodes/ComfyUI-StyleTransferPlus/models/vgg_normalised.pth

The decoder comes from the TSSAT-model.zip in the TSSAT repo's Model Testing section. The vgg_normalised.pth is the same shared file that CAST and EFDM use - if you've installed those, it's already in your models/ folder and you just need the decoder. Dependencies: torch, torchvision, scikit-image, einops.

Common issues

• Style looks too weak at default settings - that's max_iter=1 being conservative. This node is designed to be re-iterated; raise max_iter before blaming the weights.
• Load error - the decoder must sit at models/tssat/decoder_iter_160000.pth inside the pack folder, and vgg_normalised.pth at models/vgg_normalised.pth.
• Slow - inherent to the patch matching. Lower size for faster iteration tests, then raise it for the final run.

TSSAT is a "trust the process" node: leave max_iter at its default once, notice the weak result, then dial it up. It's the one in this pack where the second and third passes are the whole game.