Ultimate Affine KSampler

WASUltimateCustomAdvancedAffineNoUpscale - "Ultimate Affine KSampler" in the UI - is the pack's flagship image node, and the one behind the README's 720p → 2k demo. It takes an image, re-encodes it to latent, runs a tiled sampling pass with Affine's scale/bias enhancement baked into every few steps, and hands you back pixels. The "No Upscale" in the name is the key: unlike the sibling USDU node it doesn't scale the image up with a model first. It's a quality pass, not a size pass - low-denoise, tiled, and detail-targeted.

If you've ever done the hi-res-fix dance - upscale, re-sampler at low denoise, watch the seams - this is that idea collapsed into one node with the seams handled for you. The tiling is the Ultimate SD Upscale approach: split into overlapping tiles, sample each, blend, no VRAM ceiling on output size. The README's own "Things to take note of" on the 2k example - better rock, natural brush, preserved foreground, reduced contrast - is the Affine pitch in one screenshot: it tames the burned, over-contrasted output that speed LoRAs love to produce.

How it works

At its core it's a KSampler with three things bolted on:

Tiling. tile_width/tile_height (default 0) in image pixels. 0 disables tiling (one pass over the whole image); -1 auto-sizes tiles to ≤1024px; explicit values give you control. tile_overlap (default 64px) is how much tiles bleed into each other so the seams don't show. Tiling is what lets you process a 2K image on a card that can't touch a 2K image in one shot.

Affine, scheduled during sampling. Every affine_interval steps (default 1 = every step), the sampler applies z' = scale * z + bias through a mask generated from pattern (26 choices, default white_noise). max_scale (default 1.2) is the peak - the scale ramps as 1 + (max_scale-1)*t - and max_bias (default 0) adds a brightness offset that also scales with progress. The affine_seed controls the mask; toggle affine_seed_increment to vary it per application for temporal variety. Optionally drop a grayscale external_mask on it to gate the whole thing to specific regions, and layer the option dicts (options + pattern-specific noise_options) for finer control.

Standard sampler controls. sampler_name (default euler, 44 choices), scheduler (default simple), steps (20), cfg (8), denoise (default 0.2). That low default denoise is the whole trick - you're not regenerating, you're re-detailing. You can even pass a per-step list as cfg for schedule-shaped guidance.

The VRAM knobs are worth naming because they're why this node works on modest cards: tiled_decode (uses ComfyUI's VAE tiled decode for big outputs), batch_size (chunks the batch dimension), and merge_frames_in_batch (flattens 5D video decodes so batches concatenate safely). vae_encode_batch_mode mirrors the Wan encode node: images or frames, depending on whether your input batch is stills or video frames.

The output

One IMAGE, ready for a preview or save node - and that's the point, it's a complete replace-in-line upscale-and-refine stage.

Installing it

Part of WAS Affine (WASasquatch/was_affine). Manager: search "WAS Affine", install, restart. Or:

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

No model downloads, deps are torch/numpy (plus matplotlib). Despite the name, it explicitly does not require UltimateSDUpscale nodes.

Where people get burned

• Start with the Affine dials low. The README hammers this: latent space is sensitive, scale=0.95 can be strong. The default max_scale of 1.2 is a big swing on some models - pull it toward 1.05 before blaming the sampler.
• Model sensitivity varies wildly. The README flags that Flux/Krea-family models may want inverted (positive) scale behavior, and that Lightning LoRAs pair best with CFG 1–2. If your first run looks wrong, check these before anything else.
• Pattern choices are 26 but defaulting is fine. White noise works as a general mask; perlin/worley for organic content, velvet/ring for fine detail. Don't feel obligated to options-node every pattern.
• Tiling ≠ free. Very small tiles with heavy overlap multiply the sampling work; -1 auto-sizing is usually the best price/quality point.