[SK] Diffusion Model

This is the node you want the moment your model ships as a bare diffusion-weights file instead of an all-in-one checkpoint - which is most of the current generation. Flux, Wan, Hunyuan, Qwen-Image, Z-Image: all of them distribute the diffusion backbone separately from the text encoder and VAE, so instead of one checkpoint loader you assemble the graph from parts - this node for the model, a CLIP loader for text encoding, a VAE loader for the codec. If you've been reaching for a checkpoint loader and getting confused about why a model's download page only gives you a .safetensors with no CLIP or VAE in sight, this is the piece you were missing.

It's tucked under SK Loader/Advanced in SK-ComfyUI-FolderingLoader, a small pack from GitHub author SKFRMSEHF that reimplements ComfyUI's own loaders around a folder-tree file picker. This is the tree-picker version of ComfyUI's stock UNETLoader/"Load Diffusion Model" node - same job, same output, different combo widget.

How it works

It reads a diffusion-only weights file and loads it as a MODEL, nothing more - no CLIP, no VAE come out of this node, because the file itself doesn't contain them. The second input, weight_dtype, controls the precision the weights get cast to on load. This matters more than it sounds: for a model in the Flux weight class (12B+ parameters), the difference between loading at full precision and loading at fp8 is often the difference between fitting on your GPU and not. fp8 halves the VRAM footprint of fp16 for close to zero visible quality loss on these architectures, and cards from the RTX 40-series onward compute it natively - it's the default recipe for running large diffusion transformers on 12–24GB cards, not a compromise you're settling for.

One thing worth being precise about: this dropdown handles fp8-style precision casting on an already-downloaded file. It's not how you get GGUF quantization - GGUF files (the Q4–Q8 route for squeezing large models onto even smaller cards) are a different file format entirely and need city96's ComfyUI-GGUF loader instead. If your model's the wrong size for this node no matter what weight_dtype you pick, GGUF is the alternative path, not a setting on this node.

The inputs that matter

• unet - the diffusion-only weights file, picked from the folder-tree combo listing whatever's in your diffusion_models/ (or unet/) directory.
• weight_dtype - the load-precision dropdown. Leave it on whatever ComfyUI's default is unless you know you're VRAM-constrained; drop to an fp8 option specifically when the full-precision file won't fit your card.

Output is a single MODEL - wire it into your sampler, and get CLIP and VAE from separate loader nodes since this one deliberately doesn't provide them.

Installing it

No install section shipped with the pack, so it's the standard two routes: ComfyUI Manager → search SK-ComfyUI-FolderingLoader or the class name → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/SKFRMSEHF/comfyui_SK_Loader then restart manually. No bundled models - and note that diffusion-only weight files for current-generation models tend to run large (multi-gigabyte to 10GB+), so budget disk and download time accordingly.

Common issues

• "Where's my CLIP and VAE output?" They're not coming from here by design - this node only ever returns MODEL. Pair it with a CLIP loader and a VAE loader (this pack's SK_VAELoader, or the stock one) to get a complete pipeline.
• File fits at the wrong precision. If a model won't fit at full weights, try an fp8 weight_dtype option before assuming your card can't run it at all - that's the whole point of the setting.
• Still doesn't fit even at fp8. That's where GGUF quantization comes in, which this node doesn't do - you'd need a GGUF-format download and city96's separate loader pack for that route.
• Empty combo. The tree only lists whatever's already in your diffusion-model folder at ComfyUI startup - drop the file in and restart rather than expecting a live scan.
• No community trail. This pack has no visible discussion anywhere online, so if something's unclear, ComfyUI's own well-documented UNETLoader node behaves identically and is a fine reference (or fallback) for how this one should work.