Nodes/TEX Wrangle/TEX Wrangle
ComfyUI Node

TEX Wrangle

TEX Wrangle: per-pixel tensor expressions for images, masks, and scalars. Reference inputs with @name (e.g. @A, @base_image). Write outputs with @name = expr (e.g. @OUT, @mask, @result). Add parameter widgets with $name (e.g. f$strength = 0.5). Supports float, int, vec2/vec3/vec4, mat3/mat4, string, and array types with if/else, for/while loops, and 100+ stdlib functions. Click the ? icon for a quick reference.

By xavinitram·Created 5 months ago·Updated 22 days ago· 4
TEX Wrangle
  • _tex_any
  • in_0
  • in_1
  • in_2
  • in_3
  • in_4
  • in_5
  • in_6
  • in_7
  • in_8
  • in_9
  • in_10
  • in_11
  • in_12
  • in_13
  • in_14
  • in_15
  • out_0
  • out_1
  • out_2
  • out_3
  • out_4
  • out_5
  • out_6
  • out_7
◄code// TEX Wrangle // Read inputs with @A, @B, etc. // Write output to @OUT // Right-click → TEX Snippets for 116 examples float gray = luma(@IN); @OUT = vec3(gray); ►
◄deviceauto►
◄compile_modenone►
◄precisionfp32►
◄debug_nan_highlightfalse►
CategoryTEX

Inputs (22)

NameTypeDefaultDescription
codeSTRING// TEX Wrangle // Read inputs with @A, @B, etc. // Write output to @OUT // Right-click → TEX Snippets for 116 examples float gray = luma(@IN); @OUT = vec3(gray); TEX source code. Use @name for inputs, @name = expr for outputs. Use $name for parameter widgets.
deviceoptCOMBOautoExecution device. auto: follows input tensors. cpu/cuda: force a specific device.
compile_modeoptCOMBOnonenone: standard interpreter (recommended default). auto: EXPERIMENTAL measured auto-tier — runs the fast codegen path and trials torch.compile in the background, committing only on a measured win. torch_compile: force JIT-compile via torch.compile. NOTE: auto/torch_compile need Triton for any GPU speedup — Triton is absent on most Windows installs, where they simply fall back to the interpreter on CUDA (CPU torch.compile works but is often slower for small programs). Run `tex doctor` to see whether Triton is present. cuda_graph: CUDA-graph replay (GPU only, needs NO Triton; big win for small launch-bound programs). All tiers fall back to the interpreter on failure.
precisionoptCOMBOfp32fp32: full precision (default, recommended). auto: EXPERIMENTAL — runs fp16 ONLY where a condition-number gate proves it stays accurate (CUDA, >=1024x1024, smooth pointwise, no amplification/ill-conditioning; verified across 225 adversarial programs); everything else runs fp32. A per-cook finiteness net makes auto ~perf-NEUTRAL — it's an accuracy-safe convenience, not a speedup. fp16: EXPERT — force half-precision IMAGE temps for the raw ~1.35-1.45x win with NO safety net (coordinates & sampling stay fp32; ~1e-3 accuracy, diverges on threshold/branch/amplifying programs). LATENT stays fp32.
debug_nan_highlightoptBOOLEANfalseDBG-3: paint any pixel that is NaN or Inf bright magenta so non-finite output is visible at a glance (a 0/0, a log of a negative, an fp16 overflow). Off by default and zero-cost when off. Sits above all tiers, so it works on every execution path.
_tex_anyopt*TEX accepts any input type. Use @name in code to reference it.
in_0opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_1opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_2opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_3opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_4opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_5opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_6opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_7opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_8opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_9opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_10opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_11opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_12opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_13opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_14opt*Internal lazy input slot — mapped from wired user inputs at queue time.
in_15opt*Internal lazy input slot — mapped from wired user inputs at queue time.

Outputs (8)

NameTypeDescription
out_0*TEX output. Type auto-inferred from code.
out_1*TEX output. Type auto-inferred from code.
out_2*TEX output. Type auto-inferred from code.
out_3*TEX output. Type auto-inferred from code.
out_4*TEX output. Type auto-inferred from code.
out_5*TEX output. Type auto-inferred from code.
out_6*TEX output. Type auto-inferred from code.
out_7*TEX output. Type auto-inferred from code.