Conditioning Recast FP64

Conditioning Recast FP64 takes your conditioning - the encoded prompt that steers the model - and recasts its tensors to float64, double precision. It's a small, single-purpose node that exists because RES4LYF does a lot of its sampling math in fp64, and if you're going all-in on that high-precision path, you want your conditioning at the same precision as everything else.

This is the conditioning-side member of the pack's precision family. You've probably noticed the theme by now: ReSDPatcher defaults its style dtype to float64, LayerPatcher defaults to float64, and here's a node whose entire job is float64 conditioning. RES4LYF leans on double precision because its high-order solvers and style operations are numerically sensitive, and mixing a low-precision conditioning into a high-precision solve can throw away accuracy at the one point you were trying to preserve it. This node closes that gap.

How it works

Conditioning in, the same conditioning out with its tensors upcast to float64. It handles one required conditioning and one optional second, so you can recast a positive and negative in a single node. It doesn't change the meaning of your prompt - it changes the numeric precision it's carried in.

The inputs and outputs that matter

• cond_0 (CONDITIONING) - the conditioning to recast, usually your positive prompt.
• cond_1 (optional CONDITIONING) - a second conditioning, usually your negative, so both go through in one pass.

Outputs are cond_0_recast and cond_1_recast (both CONDITIONING) - wire them into your RES4LYF sampler in place of the originals.

How to install it

Comes with RES4LYF. ComfyUI Manager: search RES4LYF, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF/
cd RES4LYF
pip install -r requirements.txt

Portable builds use the embedded pip. Restart, hard-refresh (F5).

Common issues

Manage your expectations: this is a precision-hygiene node, not a quality slider. On a normal stock KSampler run it does nothing you'd notice - the payoff only shows up inside RES4LYF's fp64 sampling and style paths, where keeping conditioning at the same precision as the solver actually matters. Second, double precision isn't free: fp64 uses more memory and runs slower on many GPUs, so don't sprinkle it around a graph that isn't doing high-precision work. And it's not a fix for a bad prompt or a device error - it changes dtype, nothing else, so if your conditioning is producing the wrong image, the problem is upstream in how it was encoded, not in what precision it's stored at.