Affine Options

Affine Options is the one node in this family that does no work. It builds a settings dictionary - the pattern an affine masks through, plus the numbers that shape that pattern - and hands it to a sampler or to Latent Affine through an affine_options socket. Half the questions about it are really "I ran the graph and nothing changed," and the answer is almost always that the node was never wired to anything. Wire it or don't. It changes nothing on its own.

It exists because of a real problem: the affine transform multiplies a latent and adds an offset to it, and if you just do that everywhere, you cook the whole image. Distilled and flow-matching models come back a little flat - that's the tradeoff of squeezing a model into eight steps, and distillation docs are blunt about it. The community's answer to flatness for years has been Detail Daemon-style nodes that leave noise behind during the middle steps. The affine nodes do the same trick, but through a mask, which is the interesting part: grain in the sky and not on the face, texture on fabric and not on skin.

What the pattern does

The pattern widget picks one of 26 fields. Broadly:

• The noise family - white_noise, pink_noise, brown_noise, red_noise, blue_noise, violet_noise, purple_noise, green_noise, black_noise. These are grain at different coarseness. The coloured ones are spectral shapes: black_noise keeps a chosen number of frequencies alive, green_noise sits in a band, and highpass_white rolls off the low end.
• Shapes - perlin, checker, bayer, cross_hatch, tile_oriented_lines, dot_screen_jitter, poisson_blue_mask, worley_edges, ring_noise, velvet_noise. Blobs, grids, halftone screens, stipple, cell walls. Some of these are drawn flat on purpose - cross_hatch above 0.5 cycles per sample aliases into moiré, and the tooltip says so.
• Read off the latent - detail_region, smooth_region, edges_sobel, edges_laplacian. These don't exist until there's a picture to measure, which is why they're the four that a starting-noise pattern can't offer.
• solid (everything, scaled by solid_alpha) and external_mask (your own MASK, if you've wired one into the sampler).

Only the chosen pattern's own settings are drawn on the node, which is the design decision you'll either love or curse. Love: the node stays short. Curse: you tune six sliders for perlin, switch to dot_screen_jitter to compare, and your work is invisible - it's still there, it's just not on screen.

The inputs you'll actually touch

content_gate is the sleeper hit. Set it to smooth_region and a generated pattern only lands where the picture is flat, so the grain follows the subject instead of sitting over it; detail_region does the opposite, and edges_sobel/edges_laplacian pin it to edges. content_window decides how wide a neighbourhood it reads - 3 for fine texture, 15 for whole regions.

mask_strength multiplies the mask before the affine reads it, and here's the trap: 2.0 doesn't just mean "more mask", it pushes past the max_scale you set on the sampler. threshold, invert_mask, mask_blur and mask_sharpen (with sharpen_radius and sharpen_threshold) shape the mask first, then it's used. clamp with clamp_min/clamp_max is worth turning on the moment you push scale or bias hard - it keeps the latent inside numbers the model has actually seen, and it's cheap insurance.

bias_field is the other one people miss. constant adds one flat offset wherever the mask is white, which shifts colour and tone. gaussian makes it a noise field, one value per element - 0.02 is a gentle drift, 0.1 is strong.

Two groups are only for video latents and only when the sampler's temporal_mode is per_frame or drift: frame_seed_stride (how far the seed moves between frames - a big prime like 9973 makes each frame independent) and drift_speed, drift_angle_deg, drift_renew (sliding one mask across the clip). If you're running stills, they do nothing.

Installation

ComfyUI Manager, search WAS Node Suite v3, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

Requires ComfyUI 0.14.0+ and Python 3.10+. The pack installs no packages - its v3 rewrite moved everything onto torch on ComfyUI's own device. On first start it writes config.yaml under your ComfyUI user dir in was-node-suite/, and that first start takes a second or two longer.

Things that bite

If you came from a 2023-era tutorial: that WAS Node Suite is a different thing. The v2 line fought with your ComfyUI updates over pinned OpenCV and InsightFace and threw import failures after unrelated installs, and the author was open at the time about not having the hours for it. v3 installs nothing, so that whole class of breakage is gone. If you're seeing it, you have the old pack.

Otherwise: the pattern set here wins over the sampler's own pattern widget, so a stale options node quietly overrides what you typed downstream. And the node won't tell you it's working - the affine is silent unless you flip debug on the sampler.