TEX Wrangle

You know the feeling. You've dragged thirty nodes together, and then you hit the one thing no one ships a node for: a blend that's almost overlay, a mask that should be a luma key of a luma key, a procedural texture you want to hand-tune. ComfyUI's whole pitch is that every operation is a node, but the graph stops being the right tool once the operation is per-pixel math. TEX Wrangle is the escape hatch: a small programming language inside a single node, so instead of hunting for the node that doesn't exist, you write two lines of code.

It's not a texture pack, despite the name. TEX is a Tensor Expression language - the author openly cribs from Houdini's VEX, Nuke's BlinkScript, and VDB AX, which tells you exactly who this is for. If you've never written a shader, the first few minutes will be slow. If you've ever wished a pipeline tool would just let you type, it's immediately home.

How it works

You write C-ish code in the node's code field. Reference any connected input with @name - sockets are created automatically, so there's no schema to fight:

float gray = luma(@image);
@OUT = vec3(gray);

That's the whole trick. @name on the right side of = becomes an input socket; any @name you assign to becomes an output. Prefix a variable with $ and it becomes a widget on the node:

f$strength = 0.5;            // float slider
@OUT = lerp(@base, @overlay, $strength);

Underneath, the code compiles to PyTorch ops and runs vectorized across the whole image - a for loop or if/else isn't slow Python, it's tensor math. There are built-ins for pixel coordinates (ix, iy, normalized u/v), frame index (fi/fn for batch-aware video work), a 144-function stdlib for color, noise, SDFs, blends, and sampling, and static types (float, vec3, vec4, string, arrays). Right-click → Snippets dumps 116 worked examples - steal one, tweak it, run.

The inputs and outputs that matter

The code field is the whole show - everything else is optional. The few a beginner actually touches:

• device - auto (default), which follows your input tensors. Leave it alone.
• compile_mode - none is the default interpreter and the right call. auto and torch_compile need Triton to speed anything up on GPU, which most Windows installs don't have; cuda_graph is a GPU-only replay mode that's a real win for small programs. All of them fall back to the interpreter on failure, so nothing breaks - you just get the slow path.
• precision - keep fp32. fp16 exists for experts chasing a ~1.4× win and will happily diverge on threshold or branch-heavy code. auto is a nice idea (only drops to fp16 where a condition-number gate proves it safe) but it's accuracy-neutral by design, not a speedup.
• debug_nan_highlight - paints NaN/Inf pixels magenta. Turn it on the moment anything looks wrong; it's free when off.

Outputs are out_0 through out_7, with types auto-inferred from whatever you assign. Wire @OUT straight into anything that takes an image or mask - it's a standard tensor on the other end.

Installing it

ComfyUI Manager: search "TEX Wrangle" → Install, restart. Or the manual route:

cd ComfyUI/custom_nodes && git clone https://github.com/xavinitram/TEX.git TEX_Wrangle

Then restart ComfyUI. Here's the part worth celebrating in the dependency-hell era: there are no dependencies and no model downloads. PyTorch does all the work - the repo's pyproject.toml lists an empty dependencies = []. Install it, use it, and it will never be the thing that breaks your environment. There's even a standalone CLI (python -m TEX_Wrangle.tex_cli run prog.tex --in a.png --out b.png) that runs programs on image files with zero ComfyUI involved.

Common issues

• A 1×1 output with no inputs - output resolution comes from connected inputs, so purely procedural code has nothing to size itself against. Connect any image (or use iw/ih from a reference) to set the extent.
• Undefined variable 'v' or weird errors on u, ix, iy, time - these are reserved built-ins. v is the normalized y-coordinate and always defined; rename your variable to val or value.
• torch.compile doing nothing on Windows - expected. It needs Visual Studio Build Tools for CPU and pip install "triton-windows<3.7" (matching your torch version) for CUDA. The default none mode works everywhere.
• Wondering why a tier isn't engaging? Hit /tex_wrangle/doctor in your browser for a full environment report - torch/CUDA, Triton, cache state, which tiers are actually reachable.

Fair warning: this is a niche tool with a small community, so you won't find a hundred reddit threads to rescue you - you get the built-in ? card, the wiki, and the error-code docs. Because every failure is a compile error with a fix suggestion rather than a silent wrong image, that's usually enough. If you're comfortable reading an error message, TEX Wrangle quietly becomes the node you reach for whenever the graph is the wrong shape.