🦊 RS Models Loader Pro

Most ComfyUI graphs start with a sprawl of loaders - a UNET here, a CLIP there, a VAE somewhere else, and then four LoRA nodes stacked between them. RS Models Loader Pro is the RaykoStudio answer: every loader you need at the start of a workflow, collapsed into one node, with room for two of everything. It's the "just load the pipeline" hub you reach for when your workflow is getting complicated enough that you'd rather not keep four loader nodes in view.

What it combines

The Pro version is built for the graph shapes a single loader can't handle. Concretely, it can hold:

• Two diffusion models - unet_name and unet_name2. The canonical use is an SDXL base + refiner pair. It also covers the genuinely confusing case where one architecture needs two models, like Wan 2.1's separate high-noise and low-noise UNETs.
• Two CLIPs - clip_name and clip_name2, for dual-CLIP architectures (Flux, SD3, Hunyuan DiT). clip_type is the dropdown that matches the architecture: it covers flux, flux2, sd3, stable_cascade, krea2, qwen_image, wan, ltxv, ideogram4 and a couple dozen more.
• Two VAEs - vae_name and vae_name2. A classic setup: a fast preview VAE (like TAESD) for quick iterations and a proper full-quality VAE for the final decode.
• A LoRA stack - the same drag-drop, enable/disable, per-LoRA strength list as the RS LoRA Loader, with preset saving.

Every slot defaults to None on purpose. Leave what you don't need as None - the node skips it, so a simple SDXL txt2img graph only uses one of each.

The inputs that matter

• unet_name - your primary diffusion model. The one field you almost always set.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. fp8 variants halve VRAM for the model weights at a small quality cost; on a mid-range card they can be the difference between running and not.
• clip_type - must match your model family. Getting this wrong is the usual "it loaded but the output is garbage" cause.
• clip_device - default, cpu, or cuda. Forcing the CLIP to CPU can squeeze a workflow onto a small card at the cost of slower prompt encoding.
• vae_name / vae_name2 - include pixel_space as an option, handy for VAE-in-pixel-space models.

Outputs are MODEL, MODEL 2, CLIP, VAE, VAE 2 - wire the ones you used into the sampler, text encoders, and decode.

When to reach for it (and when not to)

It shines in refiner and multi-model pipelines where the alternative is two or three full loader clusters on the canvas. For a plain single-model workflow it's honestly overkill - the smaller RS Models Loader does the same job with one of everything. But if you're doing SDXL base→refiner, Wan, or preview-VAE workflows, having the whole model load in one node makes the graph dramatically easier to read and share.

Installing it

RaykoStudio is one pack, so this install gets you every RS node. In ComfyUI Manager, search for ComfyUI_RaykoStudio and install; or:

cd ComfyUI/custom_nodes
git clone https://github.com/Raykosan/ComfyUI_RaykoStudio

Restart ComfyUI after. Dependencies are real here - pycairo, opencv-python, scipy, Pillow>=10 - and Manager installs them for you; manual clones need pip install -r requirements.txt. It's Apache 2.0, tested on ComfyUI 0.15+, Python 3.10+, Torch 2.8+. Raykosan runs this pack solo and ships updates almost weekly, so when something misbehaves, check the GitHub issues first.