[Inference.Core] Content Shuffle

Content Shuffle is a strange one the first time you see what it actually produces: a scrambled, patchwork-looking mess of your source image, with the original colors and textures all present but the spatial layout deliberately scattered. That's not a bug in the preprocessor, that's the entire point. It's built for the T2I-Adapter Shuffle model, which reads "what kind of stuff was in this image, tonally and texturally" from a condition that's been stripped of composition - so your generation picks up the palette and general visual character of a reference without literally copying its layout the way Tile or Canny would.

That distinction between T2I-Adapter and ControlNet matters here, and the category this node lives in - "T2IAdapter-only" - is telling you directly: this is not a ControlNet condition. T2I-Adapters are an older, lighter-weight conditioning family, mostly an SD 1.5-era thing, and the community still points to them as the lower-VRAM alternative when a full ControlNet is more than you need. Loading a regular ControlNet checkpoint against this preprocessor's output won't work - you need the specific Shuffle T2I-Adapter model.

Inputs are all required here, unusually - most preprocessors in this pack make everything but image optional, but Shuffle requires image, resolution (default 512), and seed (default 0) all up front. That seed is the one worth understanding: the shuffle pattern itself is randomized, and the seed controls which specific scramble you get from a given source image. Same image, same seed, same shuffle - change the seed and you get a different scattering of the same underlying color/texture information, which is genuinely useful if your first attempt happened to shuffle away something you wanted preserved. The single IMAGE output feeds into a ControlNet Apply node, paired with a T2I-Adapter Shuffle checkpoint.

Practically, reach for this when you want a generation to borrow a reference's color mood and general texture - think "make this feel like that" - without any of the reference's actual composition bleeding through. It sits in a different niche from Tile (which does carry composition) and from Color, its T2I-Adapter sibling in this same pack (which carries palette but at a coarser, blockier level rather than shuffled texture).

Installing it: this comes from LykosAI's ComfyUI-Inference-Core-Nodes pack - LykosAI also makes Stability Matrix, the meta-installer a lot of people use to manage local SD UIs, so the install here is more deliberate than a typical hobby repo. Search "ComfyUI-Inference-Core-Nodes" in ComfyUI Manager, or manually: cd ComfyUI/custom_nodes && git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes, then run install.py against ComfyUI's own Python environment, or pip install -e .[cuda] / .[cuda12] / .[rocm] / .[directml] / .[cpu] matched to your hardware, or a plain pip install -e . without the acceleration extras. Restart ComfyUI once installed.

Troubleshooting: the recurring pack-wide issue is ModuleNotFoundError: No module named 'inference_core_nodes' after cloning - the Python package install didn't finish, often from install.py running against the wrong interpreter on a portable ComfyUI build (target python_embeded there specifically). Specific to this node, the most common confusion isn't a bug at all - people load a regular ControlNet model against this preprocessor's output and get nonsense results or an outright shape error, when what's actually needed is a T2I-Adapter checkpoint. If your output looks nothing like the reference's mood, try a different seed before assuming the technique doesn't work - a single unlucky shuffle can genuinely lose the character of the source.