T5v1.1 Loader

Several models in this pack - PixArt chief among them - condition on T5-XXL instead of CLIP: Google's roughly 11-billion-parameter text encoder, the same class of model behind why Flux understands natural-language prompts so much better than CLIP-only checkpoints do. It's a proper language model, not just an image-text alignment network, and that difference is the whole reason PixArt can follow a long, descriptive sentence instead of needing comma-separated keyword soup. This node is the loader for it: point it at DeepFloyd's T5v1.1-xxl weights (or a smaller converted copy) and it hands other nodes - PixArt's text-encode nodes, HunyuanDiT's text-encode nodes - a T5 object to condition on.

Why it has so many options

T5-XXL is genuinely large, and that size is exactly why this loader exposes real choices about how to load it rather than just a filename dropdown.

• t5v11_name - files or folders you've placed under ComfyUI/models/t5.
• t5v11_ver - currently just xxl, the only supported size.
• path_type - folder or file: whether you dumped DeepFloyd's four raw files loose into a folder, or you're pointing at one already-merged safetensors file.
• device - auto, cpu (the default), or gpu.
• dtype - default, auto (comfy), FP32, FP16, bnb8bit, bnb4bit, FP8 E4M3, FP8 E5M2.

Output: T5 - feeds directly into PixArt's T5 text-encode nodes or into HYDiT Text Encode (simple)'s T5 input.

Installing it

Part of the whole ComfyUI_ExtraModels pack (ComfyUI Manager: search "Extra Models for ComfyUI," or git clone https://github.com/city96/ComfyUI_ExtraModels into custom_nodes plus pip install -r requirements.txt, then restart).

For the model weights, you have real choices in file size:

• The original DeepFloyd files (config.json, both pytorch_model-0000X-of-00002.bin shards, pytorch_model.bin.index.json) into ComfyUI/models/t5, optionally in a subfolder.
• A smaller FP16-converted version - same layout, half the disk.
• A single-file BF16 merge (just model.safetensors plus config.json) - the smallest and simplest option if you don't need folder mode.

The 4-bit/8-bit modes need bitsandbytes, and on Windows you may need a newer build than whatever ships by default (pip install -U bitsandbytes). Upgrading transformers and installing spiece for the tokenizer is also required - both covered by the pack's requirements.txt.

Common issues

At full precision on CPU, T5-XXL eats roughly 22GB of system RAM. Know that before you queue a prompt on a machine that doesn't have it free - this is the kind of thing that quietly turns into your whole system swapping to death rather than a clean error.

bnb4bit mode gets VRAM usage down to around 6GB, workable on a 12GB card, but it stays permanently resident in VRAM - BitsAndBytes doesn't support the temporary CPU-offload trick normal weights get between workflows, though switching to a different workflow entirely should still release it. Older Pascal-generation cards (1080ti, P40) reportedly struggle specifically in 4-bit mode; if you're on one of those and hit odd errors, fall back to cpu. FP16 is the recommended middle ground if you've got the VRAM to spare.

If you have a second GPU in the machine, setting device to it frees your primary card's VRAM for the actual diffusion model entirely - a solid move if T5 and your checkpoint are fighting over the same 12–16GB card.