💀Prepack Model SingleCLIP

Some model families need two text encoders; a lot of the rest need just one. For those, the full DualCLIP loader is overkill - you don't want to pick two encoders when there's only one to choose. The 💀Prepack Model SingleCLIP is the trimmed-down sibling: load a diffusion UNet, a single CLIP encoder, and a VAE from one node, with a type dropdown that covers a genuinely wide swath of the ecosystem.

Its own description is the honest summary: it's the DualCLIP node's single-encoder form, and the CLIP types "follow the official Load CLIP node." Which is to say - this is the loader you'd reach for with models like Stable Diffusion 1.5/2.x, SD3-lite single-encoder setups, or any of the newer single-encoder architectures.

How it works

Mechanically it's the same sequence as the DualCLIP node, minus one encoder: it loads the UNet from your diffusion_models folder via comfy.sd.load_diffusion_model, the single CLIP from your text_encoders folder via comfy.sd.load_clip, and the VAE via the standard VAELoader, then returns all three.

The inputs:

• unet_name, vae_name - from the usual folders.
• clip_name - the one encoder.
• type - the big list: stable_diffusion, stable_cascade, sd3, stable_audio, mochi, ltxv, pixart, cosmos, lumina2, wan, hidream, chroma, ace, omnigen2, qwen_image. Each one sets how ComfyUI interprets the encoder, so pick the family your model belongs to.
• weight_dtype - default or the three FP8 options (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) for VRAM-lean loading.

The three outputs are model, clip, vae - the whole opening bundle, ready to wire into a CLIP Text Encode and a KSampler. There's also an optional device override (default vs cpu) that forces the CLIP onto the CPU, which can be a lifesaver on low-VRAM cards when the text encoder alone would blow the budget.

Where people get burned

Same type-mismatch trap as its sibling, but with a much longer menu - it's easy to select wan for a model that's actually pixart, and the node won't notice until embeddings come out wrong. The type really does need to match the model family. Also, because this node reads the CLIP type via the same mapping the official Load CLIP node uses, a brand-new architecture may not be in the list yet - the DualCLIP node's fixed four types and this one's fifteen both age with ComfyUI releases, and if your model's type isn't there, fall back to the official loader until the pack updates.

Installing it

It ships in the Prepack pack:

cd ComfyUI/custom_nodes
git clone https://github.com/S4MUEL-404/ComfyUI-Prepack.git
pip install -r ComfyUI-Prepack/requirements.txt

Or search "Prepack" in ComfyUI Manager. Dependencies: PyTorch, NumPy, Pillow. Restart, find it under 💀Prepack. No downloads in the pack - the model files go in your diffusion_models, text_encoders, and vae folders as usual.

If clip loads but embeddings come out scrambled, the type almost certainly doesn't match the encoder you picked. If the node fails to load at all, check the console for the "Failed to load CLIP model" line and verify the encoder filename is exactly what's in your text_encoders folder.