binyuan Universal Quant Converter (NVFP4/FP8/GGUF)

You've got a 24 GB checkpoint and a 16 GB card, the model is too new for anyone to have shipped a quant pack yet, and you're about to go hunt Hugging Face for a conversion that may not exist. Stop. This node is the DIY version of that scavenger hunt: point it at any safetensors sitting in ComfyUI's checkpoints, unet, or diffusion_models folders, pick NVFP4 / FP8 / GGUF, and it writes you a smaller quantized file right next to the source. No uploads, no API, nothing leaves your machine.

What it actually is

It's a converter, not a loader - a one-shot utility node living under binyuan/Advanced, named "binyuan Universal Quant Converter (NVFP4/FP8/GGUF)". You run it once, it produces a file, and then you load that file with your normal loader and go back to making images. This is a "do it yourself" tool for the long tail: the big models (Flux, Qwen-Image, Wan) get community quant packs within days, but when Krea2, ERNIE, Ideogram 4.0, or some brand-new architecture drops, a pack often doesn't exist yet. That's the gap this fills - or when you just want a specific qtype nobody bothered to upload.

How it works

The trick that makes quantized diffusion models actually load is in the details, and this node handles the parts that usually bite. It converts weights in place but keeps a blacklist of structure-sensitive layers - embeddings, norms, first/last projections, modulation, pad_token - at BF16/F32. That matters because ComfyUI infers model config from the shapes of those first/last projection layers; quantize them and the model instantiates with the wrong channels and dies on load. The remaining 2D weights (attention and FFN projections, the bulk of the bytes) get quantized.

It also writes comfy_quant markers into the output so ComfyUI recognizes the quantized layers instead of trying to load them as plain weights. If the source is already FP8, it dequantizes through weight_scale first so you don't bake in noise, and it straight-up blocks pointless conversions - like re-quantizing an FP8 file to Q8_0 GGUF, which would come out bigger than what you started with.

The three formats, and what they cost you

• NVFP4 (~0.56 bytes/param) - smallest, but the native speedup and VRAM savings are Blackwell-only (RTX 50-series + cu130 + the comfy-kitchen CUDA backend). On older cards it still runs via the full_precision_mm fallback; you just only save disk space.
• FP8 (1 byte/param) - the safest all-rounder; most newer cards handle it natively.
• GGUF - the llama.cpp ladder, with Q4_0/Q5_0 landing around 0.56–0.69 bytes/param. Caveat: the pure-Python gguf lib usually only quantizes Q8_0/Q5_0/Q4_0/BF16/F16, so K-series and IQ picks silently fall back to the nearest supported type rather than erroring.

Inputs that matter

• input_model - dropdown, auto-scanned from your model folders. The only thing you can't avoid.
• output_format - NVFP4, FP8 (e4m3fn), or GGUF.
• model_type - the blacklist strategy. Auto (Universal) works on any architecture and is what you should start with.
• gguf_qtype - GGUF only; pick Q4_0 or Q5_0 unless you have a reason.
• full_precision_mm - leave on true unless you're on Blackwell; off is where the real savings live but it'll error on hardware without the CUDA backend.
• save_to_source_dir / custom_output_dir / output_filename - where the file lands and what it's called.

Output is a single status string telling you the result or the error.

Install

ComfyUI Manager: search "Binyuan NVFP4 Converter". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yjitiu/ComfyUI-Binyuan-NVFP4-Converter.git ComfyUI_binyuan_NVFP4_Converter

Then restart ComfyUI. Two extra deps, and they're optional per format: pip install comfy-kitchen for NVFP4/FP8, pip install gguf for GGUF (usually already installed if you have ComfyUI-GGUF). If the node errors with "comfy-kitchen not detected", that's this - the node has zero declared dependencies, so ComfyUI Manager won't install them for you.

Gotchas

Where people get burned: picking a K-series GGUF qtype and getting a silently different type; trying NVFP4 with full_precision_mm=false on a non-Blackwell card; and the fake "disk full" error on Windows when your antivirus grabs the half-written multi-GB file - the node already dodges this with a temp-dir + atomic rename, but if it still trips, whitelist your models folder in Defender. Honest take: if a reputable pack already exists for your exact model, downloading it beats converting - but for anything else, this is the fastest road to a smaller model without leaving ComfyUI.