🎮 Video Anti-Aliasing (TAA + DLAA)

Let's get the one thing that will save you embarrassment out of the way: despite the name, this is not NVIDIA's DLAA. There's no closed SDK, no DLSS binary, no vendor lock-in - the README is upfront about it. VideoTAADLAA is a from-scratch PyTorch reimplementation of the temporal anti-aliasing and detail-refinement tricks modern game engines use, aimed squarely at cleaning up ComfyUI video frames. If LTX or Wan output has that crawling, shimmering edge noise on fine detail - the kind that makes hair and wires look like they're vibrating - this is the node for that exact problem.

You reach for it after generation, not during. The README's own advice is to run it as a final post-process on LTX or Wan workflows, after the VAE decode, right before the video encoder or save node. It's not an upscaler, so don't expect bigger frames: both models shipped with the pack are 1x refiners that keep the original resolution.

How it actually works

The mechanism is the classic TAA recipe, adapted to a frame batch. Each frame gets jittered by tiny subpixel offsets (it cycles through a small set of offsets, so successive frames sample slightly different spots), then a Sobel edge detector finds the jagged edges and blurs along them. The real temporal magic is in the history buffer: the node keeps the previous frame, clips it to a local variance range so it doesn't smear, and blends current against history while watching for disocclusion (motion). That blend is what kills flicker - static regions get heavily stabilized, moving ones get gated off.

Then come the detail stages: a small ~4.9MB residual CNN (DLAANet) does the "DLAA" refinement, a second texture model adds micro-detail, and a whole suite of passes handles subpixel edge reconstruction (sampling along the edge normal), specular stabilization, dehaloing, and chromatic fringe cleanup. It runs tiled at 512px with overlap, with the tile size auto-selected from your VRAM, so modest GPUs aren't excluded. A motion gate - the mean difference between current and previous frame - scales back detail boosts in moving regions to avoid ghosting.

The inputs that matter

Only a handful to touch:

• preset - Auto (picks static/balanced/motion behavior from scene motion), Performance, Balanced, or High Detail. Start with Auto. High Detail maps to the strongest internal detail preset, trading temporal smoothing for pure model-based refinement - noticeably slower.
• dlaa_intensity (0–2, default 1) - master strength for the edge/detail cleanup. This is your main quality knob.
• texture_intensity (0–2, default 1) - how hard the texture model pushes fine detail like hair and fabric.
• motion_stability (0.5–2, default 1) - crank this up if you see ghosting on movement, or down if footage is already smooth and you want sharper detail retention.

It takes an images batch and returns a single IMAGE output at the same resolution and frame count, preserving extra channels like alpha. Single still images automatically route to a photo preset with no temporal blending - so it works as a one-shot edge-cleaner too, it's just not where it shines.

Installing

The easiest install in the pack, honestly. ComfyUI Manager → search SOLRICKS (or "Anti-Aliasing Pack") → install → restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/SOLRICKS/comfyui-solricks

Then restart ComfyUI. That's the whole thing: the two model files ship inside the repo, and there are zero Python dependencies - the pack's own packaging lists dependencies = []. No model downloads, no pip installs.

Where people get burned

This is a young, small pack (single-digit stars, and you won't find much community discussion about it yet), so treat it with healthy skepticism. A couple of practical notes:

• Test on a short clip first. A handful of frames costs nothing; a 30-second render you hate costs an evening. The preset system is tuned by eye, and your content may want a different preset than the default suggests.
• Motion ghosting usually means motion_stability too low or the preset pushing temporal strength too high - dial motion_stability up, or drop to Balanced.
• Performance exists for a reason - it's noticeably faster with lighter detail cleanup, and it's the sensible pick for previews and longer videos before you commit to a final pass.
• Don't expect a resolution change. It's 1x; if you want bigger output, upscale first, then run this as the cleanup pass.

The honest verdict: for a niche post-processor that costs nothing to install and handles the single most annoying artifact in local video gen - temporal shimmer - it's a solid add. Just don't confuse it for NVIDIA's hardware path, and verify it on your own footage before you trust it on a full render.