Diffusion Safetensors -> GGUF (city96)

A ComfyUI node in Kinburg-Nodes/LLM/GGUF with 13 inputs and 3 outputs.

By Kinburg·Created 2 months ago·Updated 3 days ago· 1

gguf_path
log
help

◄model▾►

◄model_path►

◄output_dir►

◄output_name►

◄quantizenone►

◄quantize_binary▾►

◄quantize_binary_path►

◄fix_5d_tensorsauto►

◄tool_dir►

◄auto_clonetrue►

◄hf_token►

◄keep_intermediatefalse►

◄forcefalse►

CategoryKinburg-Nodes/LLM/GGUF

Inputs (13)

Name	Type	Default	Description
model	COMBO		Diffusion model from ComfyUI/models/diffusion_models or /unet. Choose the placeholder to type a path/URL in model_path. The dropdown is searchable.
model_path	STRING		Used when 'model' is the placeholder: a local .safetensors/.ckpt path, a HF file URL (…/resolve/main/model.safetensors), or 'owner/name::model.safetensors'. Give the single diffusion model file, not a folder.
output_dir	STRING		Folder for the .gguf. Created if missing. Blank -> ComfyUI/models/diffusion_models.
output_name	STRING		Base filename without extension. Blank -> derived from the source file. The precision/quant tag and .gguf are added automatically.
quantize	COMBO	none	Optional 2nd pass to shrink the model (e.g. Q4_K_S, Q8_0). none = keep the F16 gguf. Needs the PATCHED llama-quantize from city96's llama.cpp fork.
quantize_binary	COMBO		Patched llama-quantize executable (auto-listed from ComfyUI/models/llm). Choose the placeholder to type a path in quantize_binary_path. Only used when quantize != none.
quantize_binary_path	STRING		Full path to the patched llama-quantize(.exe), used when the dropdown is the placeholder.
fix_5d_tensorsopt	COMBO	auto	Wan 2.1 / HunyuanVideo need a post-quantize pass to re-add 5-D tensors. auto = run it only when the gguf's arch is wan/hyvid. Ignored when quantize = none.
tool_diropt	STRING		Path to a ComfyUI-GGUF checkout (holds tools/convert.py). Blank -> use an installed custom_nodes/ComfyUI-GGUF or auto-clone it there.
auto_cloneopt	BOOLEAN	true	If ComfyUI-GGUF isn't found, git clone it into custom_nodes. Needs git on PATH. (You'll want it installed anyway — its Unet Loader (GGUF) loads the output.)
hf_tokenopt	STRING		HuggingFace token for gated/private files. Leave blank for public models.
keep_intermediateopt	BOOLEAN	false	When quantizing, keep the intermediate F16 gguf (and the pre-fix file) as well. Off -> delete them after.
forceopt	BOOLEAN	false	Re-run even if the target .gguf exists. Off -> return the existing file without redoing the work.

Outputs (3)

Name	Type	Description
gguf_path	STRING	—
log	STRING	—
help	STRING	—