KSampler Affine

KSampler Affine is a drop-in replacement for the plain KSampler that applies the WAS Affine transform inside the sampling loop instead of as a separate node between passes. Same idea as the two-sampler manual workflow - nudge the latent where a mask says so - but the sampler handles the scheduling, so you get the effect without wiring up a second KSampler and a Latent Affine in the middle.

How it works

All the standard KSampler inputs are there - model, positive, negative, latent_image, seed, steps, cfg, sampler_name, scheduler, denoise. On top of them sit the affine controls:

• affine_interval (default 1) - apply affine every N steps. 1 = every step, 2 = every other step. Doesn't change total steps.
• max_scale / max_bias - the peak values at the schedule's strongest point. Scale is interpolated as 1 + (max_scale - 1) * t where t comes from your schedule; bias as max_bias * t.
• pattern - the mask pattern (default white_noise).
• affine_seed / affine_seed_increment - a seed separate from the sampler seed, so the mask changes independently of the image content. Increment rolls it per application, which is what you want for temporal masks.
• affine_schedule - a DICT from WASAffineScheduleOptions, interpreted over total steps. Leave it unplugged and the defaults apply.
• Optional: external_mask (gates where affine applies), options and noise_options (mask tuning DICTs), and merge_inactive_steps (greedily merges steps outside the active schedule window into bigger batches to save time; default on).

The schedule is the real control. From the author's own comparison workflow, the sensible starting shape is start 0.1 → end 0.3 with easeInOutSine: affine bites in the early-to-mid steps, then gets out of the way while the sampler settles detail. Pair it with a high-pass pattern (highpass_white, perlin) for texture, or green/pink for natural-ish detail.

Outputs: LATENT (your sampled latent, straight to VAE Decode) and MASK (the mask from the last affine application - preview it to see what got affected).

Why reach for it

It's the integrated path the README sells as "Option 2": automatic scheduling, no manual node choreography, and everything stays in latent space so it's cheap. If you're already tweaking a KSampler workflow and the image is close but the detail reads flat or the contrast looks burned (classic speed-LoRA symptom), this is a low-friction experiment: swap the KSampler, set max_scale ~1.05–1.1, and A/B it.

Installing

Part of WAS Affine by WASasquatch. ComfyUI Manager → Install Custom Nodes → search "WAS Affine" → install → restart, or:

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

Restart ComfyUI. ComfyUI 1.0.0+, torch and numpy, no model downloads.

Common issues

The usual mistake is max_scale way too high. 1.2 is the default ceiling, not a suggestion - at that strength with a hard pattern you'll reshape the image, not enhance it. Also remember the affine seed is independent: same sampler seed, different affine_seed, different result, so when you're chasing a seed, don't forget there are two of them. And if you see banding, that's the mask edges - blur them via options (blur_sigma ~1.5).