Latent Affine

The name AFFINE sounds like a lab report - "Adaptive Field Filtering by Intermittent Noise Enhancement" - but the actual idea is so simple you'll wonder why you didn't do it before. During diffusion, your image lives as a tensor in latent space. WAS Latent Affine lets you nudge that tensor, selectively, in the places where a mask says so. That's the entire pack, and this is the node that does it directly.

What it actually does

The transform is z' = scale * z + bias, applied only where a generated pattern mask is bright. Where the mask is 0, your latent is untouched; where it's 1, the values get scaled and shifted. The blend between those two extremes is what the mask controls, so the mask - not the scale - is the real star. You pick from 20+ patterns: spectral noise (white, pink, brown, blue, violet, green, black), procedural stuff (perlin, cross_hatch, dot_screen, ring_noise, highpass_white), geometric tiles (checker, bayer, solid), content-aware masks computed from the latent itself (detail_region, smooth_region, edges_sobel, edges_laplacian), or external_mask to feed your own image.

Why bother? Two big use cases. First, the classic two-pass workflow: run a KSampler for the high-noise steps, apply Latent Affine, then resume with a second sampler for the low-noise tail. It's the latent-space cousin of a hi-res fix second pass - same "adjust between passes" instinct, but you never leave latent space, so no VAE decode/encode round trip and none of the raster-space artifacts. Second, quality salvage: the README frames it as rescuing detail in light LoRA outputs and taming the burned, over-contrasted look that speed-boosting LoRAs (Lightning/Turbo) love to produce.

The inputs that matter

You'll set three things 95% of the time, everything else is optional tuning:

• scale - multiplicative factor. Default 0.96. Below 1 darkens, above 1 amplifies features. "0.95 can be quite strong" is the author's own warning.
• bias - additive offset, -2 to 2. 0.1 brightens, -0.1 darkens.
• pattern - the mask source. Start with perlin (organic) or white_noise and get fancy later.

Then temporal_mode (static = one mask for all frames, per_frame = regenerate each frame for video), seed, and the two option inputs: options (the shared mask-tuning DICT from Latent Affine Common Options or the Super Options node) and noise_options (a per-pattern DICT that overrides the base options). If you connect an external_mask while using any procedural pattern, the generated mask gets multiplied by it - gating the effect to specific regions without giving up the pattern.

Outputs: latent (wire into your next sampler or straight to VAE Decode) and mask (the generated mask - worth previewing once so you actually see what's being affected).

Installing

This is a node in WAS Affine by WASasquatch - the same author as the legendary WAS Node Suite. In ComfyUI Manager, hit Install Custom Nodes, search "WAS Affine", install, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was_affine

Restart ComfyUI after. You need ComfyUI 1.0.0+, and the dependencies are just torch and numpy - no model downloads, no extra checkpoints. Despite the "USDU" names elsewhere in the pack, nothing here needs UltimateSDUpscale.

Where people get burned

Latent space is sensitive, and it's the whole selling point, so respect it. Start with scale near 0.95 and a soft pattern, not 1.2 and a hard checker. Some models respond far more dramatically than others - the README notes Flux/Krea-family models can show inverted behavior and may want positive scale values instead. If you see banding at mask edges, that's the mask, not the model: blur it (blur_sigma ~1.5 via options) before it ever touches the latent. And if per_frame makes video look like static noise crawl, dial the effect down before blaming the temporal mode - that liveliness is literally the point of the setting.