DualCLIPLoader (Dynamic VRAM)

Most of the models you'll actually run don't use one text encoder - they use two. Flux pairs a giant T5-XXL with a CLIP-L. SDXL uses two CLIPs. SD3.5, Hunyuan, LTX, and the newer Krea 2 stack all want a pair. DualCLIPLoader (Dynamic VRAM) is the version of the GGUF CLIP loader that handles two files at once, and it's the one you'll reach for whenever a workflow says "DualCLIPLoader" but you want the quantized files.

The pitch is the same as the single-file version: quantized text encoders free VRAM the checkpoint can use. T5-XXL at Q8 is roughly half its 4.7GB fp16 footprint, and on a 12GB card that headroom is real. Combined with a GGUF checkpoint, it's the difference between a Flux workflow that fits and one that doesn't.

How it works

It's the GGUF loader's load_patcher path with two files. Both filenames come from the same combined dropdown (models/clip + models/text_encoders), and both GGUF and regular safetensors files are accepted - mix a quantized T5 GGUF with a plain CLIP-L safetensors if that's what you have. The type dropdown is the important bit: it tells the loader which architecture combination you're building, and it's populated from ComfyUI's own DualCLIPLoader types - flux, sdxl, sd3, ltxv, wan, hidream, hunyuan_video, and more. The Dynamic VRAM wrapper then hands both encoders to the aimdo memory manager so they load, offload, and reload sensibly.

The inputs that matter

• clip_name1 and clip_name2 - the two encoder files.
• type - flux, sdxl, sd3, ltxv, wan, hidream, etc. Match it to the model.

Output is CLIP, straight into your CLIP Text Encode nodes.

Installing

Search ComfyUI-GGUF in the ComfyUI Manager, or:

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

Restart ComfyUI and pip install --upgrade gguf. ComfyUI v0.27.0+ for the native INT8 path; some newer encoders (MiniMax H3 pruned, Gemma 4) want newer ComfyUI builds still.

Where people get burned

• The Dynamic VRAM variants need a supporting ComfyUI - launch without --disable-dynamic-vram, or the loader errors out with a message telling you exactly that.
• Mix scaled-FP8 and GGUF at your own risk - it's rejected. The single-file loader refuses that combination, and the multi-file path inherits the same rule: keep quantized encoders on the GGUF side.
• Pick the right type. flux vs sd3 vs ltxv are not interchangeable; the embeddings come out wrong if the architecture doesn't match the checkpoint.
• The _K-quants rule only applies to diffusion models. Text encoders are the one place this fork does support Q4_K_M-style quants, so a quantized T5 is a legitimate move here.