TENSOR Booster

Don't let the name sell you on more than this node actually does. TENSOR Booster takes a safetensors file and re-saves it at fp32 precision - but "boosting" a bf16 or fp8 file up to fp32 doesn't hand you back detail that was already thrown away. It's a container upgrade, not a quality upgrade. What it's actually for is the specific situation where a downstream tool insists on fp32 input, most commonly a VAE.

Why this exists

The pack's own README explains the real use case, and it's tucked inside the documentation for a different node: their TENSOR Cutter, which halves precision from bf16 to fp8, comes with a warning that certain tensors - VAE weights especially - often need to stay at fp32 to work correctly, and cutting them more aggressively can produce a black-screen result. The fix the author suggests is to boost the source to fp32 first (if you don't have the original fp32 file anymore), then cut from that fp32 baseline. This node is that boost step.

How it works

It runs standalone - no wiring in, no wiring out. Select a safetensors file, hit Queue, and the console shows progress while it re-encodes the tensors at fp32. The output lands in ComfyUI/output when it's done.

Inputs and outputs

One input: select_safetensors, a dropdown of the safetensors files it can see. No outputs in the graph sense - this is a true output node (is_output_node: true): the result is a file written to disk, and nothing downstream depends on it.

Installing it

ComfyUI Manager: search gguf, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/calcuis/gguf

No extra pip installs on the current version. Put the file you want to boost into ComfyUI/models/diffusion_models so it appears in the dropdown.

The honest take

If you're expecting this to fix a model that already looks worse from a previous quantization pass, it won't - precision that's already gone doesn't come back by re-representing the same numbers in a bigger container. The only legitimate reason to run this is when something downstream - most often a VAE, per the pack's own documentation - specifically needs fp32 to behave correctly, and you don't have a genuine fp32 source file to hand it. Running this on every model "just in case" only costs you disk space for no benefit.

Common issues

Expecting a quality improvement. There isn't one. If your goal is better output, this is the wrong tool - look at your sampler settings, your quant level, or the model itself instead.

Running it on files that don't need it. Most tensors don't have the fp32-or-bust requirement that VAE weights sometimes do. Reach for this specifically when you've hit the black-screen-from-cutting problem this pack's README flags, or another tool explicitly demands fp32 input - not as a routine step.

No community troubleshooting trail. This is a low-traffic utility node with essentially nothing written about it beyond the pack's own README - there's no body of known failure modes beyond what's documented here to lean on if something behaves unexpectedly.