TENSOR Cutter (Beta)

fp8 is the community's default first move before anyone reaches for GGUF at all - near-zero quality loss, half the VRAM of full precision, and it's what most people should try before quantizing further. TENSOR Cutter produces exactly that: it takes a bf16 safetensors file and cuts it down to fp8, locally, without needing to find someone else's pre-cut version online. It's still labeled "Beta" in the node menu, which is the author telling you directly not to treat it as bulletproof yet.

Why you'd want to cut your own fp8

Popular models usually already have community fp8 releases floating around. Less popular ones, fine-tunes, or anything you trained yourself often don't. This node fills that gap - if you've got a bf16 file and want the fp8 version without waiting for someone else to publish one, this cuts it for you directly in ComfyUI.

How it works

Standalone, same as this pack's other conversion tools - nothing wired in or out. Select the safetensors file, hit Queue, watch the console for progress. When it's done, the cut fp8 file lands in ComfyUI/output.

Inputs and outputs

One input: select_safetensors, a dropdown of the safetensors files it finds. No graph output - this is a true output node (is_output_node: true), producing a file on disk rather than a value for another node to consume.

Installing it

ComfyUI Manager: search gguf, install, restart. Manual:

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

No extra dependencies to install on the current version. Drop the file you want cut into ComfyUI/models/diffusion_models so it shows up in the dropdown.

The one real gotcha

The pack's own README flags this directly, and it's worth taking seriously: some tensors - VAE weights being the common example - need to stay at fp32 to function correctly. Cut one of those too aggressively and you can end up with broken output, black screens being the specific symptom the author calls out. If that happens to something you've cut, the fix isn't to fight the fp8 version - go back to the original file (or restore it to fp32 first with this pack's TENSOR Booster, if you no longer have a genuine fp32 source) and cut from a clean baseline instead.

Common issues

Black screen or broken output after cutting. This is the documented failure mode - you likely cut a tensor (often VAE-related) that needed to stay at higher precision. Restore from fp32 and try again rather than assuming the cutter itself is broken.

It's still "Beta." The author hasn't called this stable. Keep your original bf16 file around rather than treating the cut version as a replacement - this is a "try it, verify it works, then decide" tool, not a "trust it blindly" one.

Quality loss you didn't expect. fp8 is generally the safe, near-lossless step - but "generally" isn't "always," and results can vary by model. Compare output against the original before committing to the cut version for anything you care about getting right.

No community reports to lean on. Like this pack's other conversion nodes, there's essentially no public discussion of this one specifically - you're working from the README and direct testing, not a body of known community fixes.