🦊 RS Models Loader

A checkpoint is really four things - a diffusion model (UNET), a text encoder (CLIP), a VAE, and a stack of LoRAs - and normally that's four or five separate nodes sprawled across the graph. 🦊 RS Models Loader (RaykoModelsLoader) collapses them into one box. Wire its MODEL, CLIP and VAE outputs to your sampler, and everything from loading to LoRA stacking happens in a single interface. If you're swapping between model configs all day, that's a meaningful chunk of your graph that just got smaller.

How it works

Under the hood it's not reinventing anything - the source calls ComfyUI's own UNETLoader, CLIPLoader / DualCLIPLoader, VAELoader and LoraLoader nodes. What it adds is a single entry point plus two things those native nodes don't give you: per-node presets that save an entire model set (UNET + CLIP + VAE, without touching your LoRA list), and an inline LoRA manager where you can add, toggle, and set independent model/CLIP strengths.

The dual-CLIP handling is the part worth calling out. For Flux, SD3, Hunyuan DiT and PixArt you flip use_clip2 on and both CLIP slots go live; leave it off and clip_name2 is visible but greyed out. It auto-falls back to single-CLIP mode if clip_name2 ends up empty, so you won't hard-crash on a mismatch.

Inputs that matter

• unet_name - the diffusion model. Tooltip: "Diffusion model (UNET)".
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. If your card is tight, fp8 gets you running on a lot less VRAM.
• use_clip2, clip_name, clip_name2 - single vs dual CLIP. clip_type picks the architecture (27 options covering stable_diffusion, sd3, flux, flux2, qwen_image, krea2, ltxv, wan and more).
• clip_device - default, cpu, or cuda. Sticking CLIP on CPU frees VRAM for sampling.
• vae_name - the VAE, including a pixel_space option for direct pixel manipulation.
• lora_data - a JSON-ish string holding your LoRA stack; the UI builds it for you.

Outputs: MODEL, CLIP, VAE - plug them straight into KSampler, CLIPTextEncode, and VAEEncode/VAEDecode.

How to install

It's part of ComfyUI_RaykoStudio:

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

Then restart ComfyUI. Or install via ComfyUI Manager by searching "ComfyUI_RaykoStudio". It needs your models to live in the standard ComfyUI folders (unet, clip, vae, loras) - the node uses the same folder paths as the native loaders, so whatever you've already downloaded is pickable.

Common issues

• LoRA config lost between sessions? It shouldn't be - the README says LoRA settings persist per node and save with the workflow. If they're not sticking, it's usually because you closed the browser tab before the workflow was written back.
• Wrong clip type. If your CLIP output produces garbage or errors, double-check clip_type matches the model. Flux needs flux, not stable_diffusion - the dropdown looks the same but the encoders are completely different.

Honest take: if you run one model and never touch it, this is unnecessary weight. If you juggle Flux, SDXL and a Qwen-Image setup in the same week, the preset system alone is worth the install.