TaylorSeer

The name sounds like an API you'd need a key for, but TaylorSeer is none of that. It's a node that takes your existing Flux or HiDream checkpoint, patches it in memory, and makes it render roughly twice as fast - no training, no distillation LoRA, no extra download. You feed it a MODEL, it hands you a faster MODEL, and you keep your whole workflow as-is. If you've been eyeing TeaCache and wishing it didn't blur the composition, this is the more interesting sibling.

What it actually does

TaylorSeer belongs to the "training-free caching" family (TeaCache, ToCa, Delta-DiT). The idea: most denoising steps are redundant, so instead of running the whole transformer every step, you run it fully on some steps, cache the intermediate activations, and cheaply predict them on the steps in between.

What's different here is the prediction. Rather than just replaying a cached tensor (TeaCache's blunt approach) TaylorSeer treats the modulation gates inside each DiT block as Taylor-series derivatives and fits an actual polynomial to each feature over time. On a skipped step it evaluates that polynomial - sum (1/i!) * f^(i) * x^i with order up to max_order - instead of recomputing. It's genuinely clever math, and it's why the author's comparison vs. TeaCache shows more consistent composition at the same speedup. There's a cost, though: unlike TeaCache, the cache lives in VRAM, and the standard version is not free memory.

The settings that matter

The node is thin on the surface - one MODEL in, one MODEL out, wire the output straight into your sampler. The five knobs are:

• model_type - flux or hidream. The Taylor cache is patched per-architecture, so this must match your checkpoint or nothing happens.
• first_enhance - the big one. It's the number of early steps that always compute in full. The README says first_enhance = 10 at 30 steps is "almost lossless" at 2x. Community testing agrees: below 10, quality falls off a cliff. This is the dial to protect.
• last_enhance - same idea for the end of the schedule (default 50). The high-detail tail stays honest.
• fresh_threshold - how many steps the cache can ride before a forced full recompute (default 6). Lower = better quality, less speedup.
• max_order - Taylor polynomial order (0–2). This is also your VRAM bill: order 0/1/2 cost roughly +2/4/6GB on Flux fp8 at 1024², and +5/10/15GB on HiDream. If that makes you wince, use TaylorSeerLite instead.

The optional block_swap_args input exists so you can add a FluxBlockSwap or HidreamBlockSwap node for VRAM relief - more on those separately.

The honest speedup

The README's headline numbers are real but optimistic. In practice the community consistently lands around 1.7–2x with settings that keep quality intact (e.g. first_enhance=10, fresh_threshold=6, 30 steps). The 3–5x claims require much more aggressive thresholds, and at that point you're trading real detail for speed - the same trade every cache method makes, just with less smearing than TeaCache. It's a 2x-ish node. That's still a great deal for a zero-download patch.

Install

Nothing exotic here, and that's the point - no requirements.txt, no model files:

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

Restart ComfyUI, or search "ComfyUI-TaylorSeer" in Manager. The one real requirement is ComfyUI itself: the README insists you be newer than commit c496e53. This pack monkeypatches ComfyUI's internal Flux/HiDream forwards, so it's inherently version-sensitive - several people have reported it silently doing nothing after a ComfyUI update. If that happens, update both ComfyUI and the pack; the repo is still getting commits (Flux 2 and Z-Image support landed in early 2026).

Where people get burned

• VRAM. The standard version's cache genuinely costs gigabytes. If your card is already near the ceiling, skip straight to TaylorSeerLite - nearly zero VRAM, same idea.
• first_enhance too low. Crank it to 10+ before blaming quality.
• Wrong model_type. Flux checkpoint with hidream selected and you'll get garbage or an error, not a speedup.
• Expecting 5x. Set expectations at 2x and you'll be pleasantly surprised instead of disappointed.