📥 QuadrupleCLIP Download

Some of the newer models don't run on a single text encoder anymore. HiDream is the poster child: it wants four of them stacked together - clip_l, clip_g, a T5, and a Llama-based LLM encoder - before it'll turn your prompt into anything. This node is the loader for exactly that case. Four dropdown slots, pick an encoder for each, and it hands the diffusion model one combined CLIP bundle. If a file you pick isn't on disk yet, the SDVN download plumbing grabs it for you instead of erroring out.

Why four encoders

Worth understanding what you're actually loading. The text encoder is a separate network that runs before the diffusion model and converts your prompt into vectors the model cross-attends to - it's not part of the checkpoint, it's its own file. SD 1.5 used one CLIP. SDXL used two. Flux and SD3 bolted a T5 on top. The current wave of models leans on a general-purpose LLM as one of the encoders, and HiDream in particular expects all four working together. Miss one, or load the wrong one, and you don't get a subtle quality dip - you get noise or a hard error. So a node that loads all four in the right shape is genuinely useful, because wiring four separate loaders by hand is tedious and easy to botch.

The inputs and outputs that matter

There are four inputs and they're all the same kind of thing:

• CLIP_name1–CLIP_name4 - one dropdown each, listing the curated encoder files (clip_l, clip_g, the T5 variants, the Llama/LLaVA encoders, and so on). You're filling four slots with the four encoders your model expects. Order matters - match what the model documentation asks for.

The single output is CLIP, which wires straight into your text-encode node (a CLIP Text Encode) exactly like any other CLIP source. From there it feeds conditioning into your sampler.

That's the whole node. No strength dials, no URL box - the download happens automatically when a selected file is missing, using the same civitai/huggingface fetch the rest of the SDVN Download family uses.

Installing it

ComfyUI Manager is the easy path: search SDVN_Comfy_node, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node

Then from your ComfyUI root install the deps and restart:

pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt

On Windows or macOS you also need to install aria2c yourself - that's the accelerated downloader the auto-fetch relies on. Linux usually already has it.

Where people get burned

The big one is VRAM. That T5 and the Llama encoder are large files, and the encoders now make up a serious chunk of a model's total footprint. If you're tight on memory, reach for the fp8 variants of the heavy encoders - fp8 is half the size of fp16 with near-invisible quality loss for inference, and the community's standing advice is basically "if it fits at fp16 use fp16, otherwise just use fp8 and don't worry about it." You'll see both in the dropdowns.

Second: this node assumes the four-encoder architecture. If your model only needs one or two encoders, you're using the wrong loader - grab the CLIP or DualCLIP download node instead. Feeding four encoders to a model that wants two won't magically work.

And the usual pack-wide note: if a file fails to auto-download, it's almost always the missing aria2c on Windows/macOS, or a bad URL for a gated repo. Check the ComfyUI console - the SDVN nodes log what they're fetching.