VOID Load Text Encoder

VOID Load Text Encoder loads the T5-XXL text encoder that turns your prompt into the conditioning VOID's video transformer reads. The pleasant surprise: it's the same t5xxl_fp8_e4m3fn.safetensors file every FLUX setup already has in models/clip/, so if you've ever run Flux you probably don't need to download anything.

The node exists because the VOID pipeline needs the raw HuggingFace T5EncoderModel and T5Tokenizer, not ComfyUI's CLIP abstraction. ComfyUI's built-in Load CLIP wraps text encoders in its own conditioning machinery, and that wrapper is incompatible with how the pipeline here assembles the diffusers-based CogVideoX inpainting model. So this pack gives you a loader that returns a custom VOID_TEXT_ENCODER type, and only VOID Loader knows what to do with it.

The one input

t5_name is a dropdown over models/clip/. It filters to filenames containing "t5" - in practice that means your t5xxl_fp8_e4m3fn.safetensors. If it can't find any t5-named file it lists everything in the folder so you can still pick manually.

The output VOID_TEXT_ENCODER carries both the encoder and its tokenizer, so one wire into VOID Loader is enough.

fp8 is fine, that's the point

The README's recommended file is the fp8 one (t5xxl_fp8_e4m3fn.safetensors, ~5 GB, from Comfy-Org/flux1-dev). Don't chase down a full-precision T5: this node casts fp8 weights to bf16 on load, and fp8-for-T5 is one of those rare quantizations where the quality cost is close to invisible for most prompts. On a 12GB card the fp8 T5 is basically the difference between the whole workflow fitting and not.

One auto-download to know about: the T5 tokenizer (~800 KB) is pulled from HuggingFace on first use and cached inside the pack's configs/t5_tokenizer/ folder. It's a one-time thing and needs internet on the first run - after that it uses local_files_only, so no network calls.

Getting set up

Install the pack (ComfyUI Manager → search "VOID", or git clone https://github.com/shanef3d/ComfyUI-VOID into custom_nodes/ and restart). Drop the fp8 T5 file into ComfyUI/models/clip/ if it isn't there already, refresh the node list, and the dropdown will find it.

The usual failure is the dropdown being empty or the node throwing FileNotFoundError with the expected path and source URL printed. That means the file isn't in models/clip/ under the name you selected - check the folder. If you're on a machine that's never seen FLUX, this is also the node that forces you to face the pack's dependency requirement: diffusers>=0.33.1, newer than stock ComfyUI ships, and the pack hard-checks it at import. Upgrade diffusers with pip install "diffusers>=0.33.1" and be ready to pin any other custom nodes that pin diffusers themselves - they conflict by design, not by accident.

There's genuinely not much else to it. It's a loader for a file you own, wired to one consumer. Set it, forget it, and move on to the part of the workflow that actually does the work.