CLIPLoader (Dynamic VRAM)

Everyone remembers to quantize the big transformer. Almost nobody remembers that the text encoder eats VRAM too - and on Flux-class workflows the T5-XXL encoder is a 4.7GB fp16 slab sitting next to a 12B model. CLIPLoader (Dynamic VRAM) is the node that shrinks that slab. It loads quantized text encoders (T5-XXL GGUF, Qwen3-VL, Gemma 4, MiniMax H3 pruned - whatever fits your model) and hands you a CLIP object exactly like the built-in loader, just lighter.

The win is cumulative. Run a Q4-Q8 checkpoint and a quantized T5 and you've freed enough VRAM to bump up a resolution or stop offloading entirely. This is the move that gets Flux-class stacks onto 8–12GB cards, and the community's been doing it since the format landed.

How it works

This is a text-encoder loader from the molbal fork of ComfyUI-GGUF, so it's built on city96's original CLIP GGUF loader. It reads a single file from your models/clip or models/text_encoders folder. Here's the part people miss: it handles both .gguf and regular .safetensors files - the dropdown lists both, so you can mix and match without changing nodes. One real restriction: it won't mix scaled-FP8 files with GGUF ones; that combination throws a NotImplementedError, so keep quantized encoders on the GGUF side.

The "Dynamic VRAM" variant wraps the loaded encoder in the aimdo memory manager's patcher, so the encoder's weights are managed the same smart way as the checkpoint. You pick the encoder type from the type dropdown - stable_diffusion, sd3, ltxv, wan, hidream, chroma, and two dozen more - matching the architecture you loaded, same as the core CLIPLoader.

The inputs that matter

• clip_name - which file to load, from models/clip (or models/text_encoders). Both GGUF and safetensors appear here.
• type - the encoder architecture: stable_diffusion, sd3, ltxv, wan, hidream, chroma, and more. Get this wrong and the encoder won't produce usable embeddings.

Output is CLIP, which goes into your CLIP Text Encode node. If your model needs two encoders (Flux = T5 + CLIP-L, SDXL = two CLIPs), that's what the Dual/Triple/Quadruple variants are for - this single-file version is for models with one encoder, or for loading encoders one at a time.

Installing

Same as every node in this pack: search ComfyUI-GGUF in the ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/molbal/ComfyUI-GGUF

Restart, then pip install --upgrade gguf. Requires ComfyUI v0.27.0+. Pre-quantized encoders live on HuggingFace - city96's t5-v1_1-xxl-encoder-gguf is the usual starting point.

Where people get burned

• The Dynamic VRAM nodes need a ComfyUI build with the aimdo memory manager and a launch without --disable-dynamic-vram, or they refuse to load.
• For newer encoders, your ComfyUI build matters in specific ways - MiniMax H3 pruned encoders want a build with MiniMax support, and Gemma 4 GGUFs need v0.30.0+. If the dropdown's type list is missing the architecture you need, update ComfyUI first.
• _K quants are supported for text encoders (unlike diffusion models), so Q4_K_M T5s are fair game here - just don't expect the same on the checkpoint side.