Diffusion Model Dtype Conversion

Sometimes you don't want a different model, you want the same model at a different precision. An fp32 UNet is twice the size of its fp16 version; a bf16 version is the size that fits. This node does that conversion for standalone diffusion models in models/diffusion_models, streaming tensor by tensor so a multi-gigabyte file doesn't need to load into RAM all at once. It's the precision resize tool for the "this model won't fit / this model is too big to merge comfortably" problem.

How it works

Pick a file, pick a target dtype (fp32, fp16, or bf16), give the output a name, run it. The node streams the safetensors through the pack's Unified Efficient Loader, converts every floating tensor to the target dtype, preserves all non-floating tensors exactly as-is, and writes the result atomically to models/diffusion_models. Because it streams, it works on models far larger than your RAM - the memory footprint stays bounded by a single work unit.

The interesting control is exclude_patterns. It's a multiline field of Python regexes, one per line; any tensor key matching a pattern keeps its original dtype instead of converting. This exists because not every floating tensor wants to live at reduced precision - the classic case is small bookkeeping tensors like norms and scalers, or a conditioning embedding you want to keep in fp32 for precision. The regex gets a clear error if you feed it something unparseable, and an invalid key pattern fails loudly on the line number, so debugging a typo is painless.

One hard rule in the code: the output path must differ from the input path. You can't convert a file over itself, which is a good guardrail - you'd rather end up with two files than a corrupted one.

Inputs that matter

• model_name - the diffusion-model file to convert.
• target_dtype - fp32, fp16, or bf16.
• exclude_patterns - regex lines for tensors that keep their original dtype. Leave empty to convert everything floating.
• output_filename (default converted_model) - no extension, written to models/diffusion_models.

Outputs

output_path (the written file - the node is an output/terminal node, so it just reports where the file landed) and conversion_report (a text summary of what was converted and what was preserved).

Why you'd reach for it

Two realistic reasons. Downcast: you have an fp32 diffusion model and you want the fp16/bf16 version without hunting for a re-upload - this halves the file and the VRAM cost at negligible quality cost for most content. Upcast: you're about to feed the file into a merge, and CWB does its arithmetic in fp32, so converting a model that's been living in bf16 up to fp32 first can avoid precision surprises in the merged result. The exclude patterns are what make it safe for the first case - keep the fragile tensors high-precision and the bulk at half precision.

Install

Part of Model Utility Toolkit (silveroxides/ComfyUI-ModelUtils). ComfyUI Manager → search "Model Utility Toolkit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

Restart ComfyUI. Real dependency: unifiedefficientloader (UEL) for the streaming loads. Keep ComfyUI current - the pack uses the newer extension API.