Load StableMaterials

Every workflow in comfyui-stable-materials starts here, and not just out of habit: the two generate nodes don't hide a model behind a convenience widget the way a typical ComfyUI pack does. They demand a pipeline input, and this node is the only thing that makes one. Load StableMaterials is the pack's entry point - it pulls the actual diffusion model into VRAM and hands you a handle you feed to Generate PBR (Text) or Generate PBR (Image).

The model behind it is StableMaterials by Giuseppe Vecchio, loaded through HuggingFace's diffusers library. It's a diffusion model trained on the MatSynth dataset that outputs five physically-based rendering (PBR) maps at once: basecolor, normal, height, roughness, and metallic. If you've come from the image-to-3D side of ComfyUI, PBR is the same concept that makes Hunyuan3D 2.1 and TRELLIS.2 textures relightable - the material ships as separate channels a game engine or renderer can light, not one baked image.

One input, two modes

The entire config surface is a single enum, mode:

• LCM (4 steps, fast) - the default you want for iterating. Roughly 2 seconds per generation on a 4090.
• Standard (50 steps, quality) - the full sampling schedule when you've settled on a material and want the best result.

Reading the source, the mode switch is doing real work. In LCM mode it loads the unet_lcm subfolder of the model repo and swaps in an LCMScheduler; in standard mode it loads the full pipeline. Both use trust_remote_code=True because StableMaterials ships custom pipeline code, not just weights.

The catch nobody mentions: the cache

The pipeline is cached globally per mode. That's great - generate once, reuse cheaply. But it means the first run in a session does the heavy lifting: ComfyUI reaches out to HuggingFace, downloads the model, and loads it. The README doesn't shout about this, so plan for a long first generation.

The sneaky part: if you switch mode on the loader, the old pipeline is deleted and torch.cuda.empty_cache() is called before the new one loads. Expect a VRAM spike and a pause. If you run both an LCM and a standard graph in one session, you'll pay the load cost twice. My advice: pick a mode and stick with it for a session.

The output is a single pipeline of type SM_PIPELINE - it doesn't do anything on its own, it just feeds the generate nodes.

Install

The README recommends ComfyUI Manager: search for "StableMaterials" and install. Manual install is the usual two steps:

cd ComfyUI/custom_nodes
git clone https://github.com/P5ina/comfyui-stable-materials.git
cd comfyui-stable-materials
pip install -r requirements.txt

Then restart ComfyUI. Requirements are thin: diffusers, transformers, and accelerate - nothing exotic, which is refreshing. The model itself downloads to your HuggingFace cache on first use (a few GB), not to models/, and it needs HF reachable. The repo is MIT; the model is OpenRAIL, so commercial use of generated materials is fine.

Where people get burned

Because the model uses trust_remote_code, an old or mismatched diffusers version can throw on load - if the pipeline errors at the loader, update the three requirements first. And if generation is slow on your first try, check you're actually in LCM mode; standard mode at 50 steps is meant to be slow.