TaylorSeerLite

Here's the honest thing about the standard TaylorSeer node: its cache costs gigabytes of VRAM, and on a 12GB card that's a dealbreaker. TaylorSeerLite is the fix the author shipped in September 2025 - the same Taylor-cache acceleration, but it keeps far less state and claims "nearly zero" VRAM increase while it's at it. It's also the only node in this pack that handles the newer model zoo: Flux 2, Wan 2.1/2.2, Qwen-Image, Lumina 2, and Z-Image/Flux.2. If you're on a consumer card, this is the TaylorSeer you actually want.

What it is and how it works

Same core trick as the full version: run the transformer fully on selected steps, cache the intermediate activations, then forecast the skipped steps with a Taylor polynomial built from the model's own modulation gates. The difference is architectural - Lite patches the model's single top-level forward pass rather than every individual block, which means the cache bookkeeping is a fraction of the size. That's where the "nearly zero VRAM" claim comes from, and it's why Lite is the version that made Wan video acceleration practical: the README quotes 386s vs 1176s (3.05x) for 81 frames on an RTX 5090.

The input list looks identical to TaylorSeer, but the model_type enum is the headline:

• flux, flux2, hidream, wanvideo, qwenimage, lumina2

Everything else carries over: first_enhance (the quality dial - keep it at 10+), last_enhance (default 50), fresh_threshold (refresh interval, default 6), max_order (Taylor order, default 2), plus the optional block_swap_args for the matching BlockSwap node. Output is a MODEL that feeds your sampler, exactly like the standard node.

One subtlety for the two-pass crowd: on Qwen-Image and Lumina 2, Lite keeps a separate cache for the negative/uncond pass so the two don't clobber each other. It's handled internally - you don't set anything - but it's why quality holds up on models that use CFG.

The honest expectations

The community's read on TaylorSeer applies here too: the realistic win is ~2x, sometimes 3x on video, with near-lossless quality at first_enhance=10 over 30 steps. Push the thresholds hard and you'll see bigger numbers with visible degradation - same law as every cache method. And yes, the "TaylorSeer doesn't work on Flux 2" chatter you might find is from before the pack added flux2 support; the current repo ships working Flux-2 examples (plus Z-Image).

Install

No dependencies beyond ComfyUI itself - no requirements.txt, no model files to fetch:

cd ComfyUI/custom_nodes
git clone https://github.com/philipy1219/ComfyUI-TaylorSeer

Restart, or install via Manager by searching "ComfyUI-TaylorSeer". Keep ComfyUI newer than commit c496e53; this pack hooks ComfyUI's model internals, and people have reported runs silently no-op'ing after a big ComfyUI update. Update both sides and it comes back.

Troubleshooting

• VRAM still tight? You're probably on standard TaylorSeer by mistake - make sure the node is TaylorSeerLite, not TaylorSeer. That's the whole point of the name.
• Wan runs slow regardless - SageAttention and torch.compile help video a lot; Lite is the acceleration, it isn't a substitute for memory-efficient attention.
• Quality dipped - first_enhance below 10 is the usual culprit, or fresh_threshold is too high for your step count.
• Nothing seems faster - confirm model_type matches your checkpoint. A mismatched type either errors out or, worse, runs uncached.