Emu 3.5 Loader

Every Emu 3.5 workflow starts the same way: you need the model, its tokenizer, and the vision tokenizer loaded and handed to a sampler. That's this node's entire job - Emu 3.5 Loader is the V1 gateway, the original way to get BAAI's model into a ComfyUI graph. It's simple, it's a bit old now, and it still works fine.

It lists whatever folders it finds inside ComfyUI/models/emu35/ and pops them into two dropdowns - one for the main model, one for the vision tokenizer. On load it reads config.json to decide how to proceed, and the precision dropdown gives you the strategy:

• auto - detects whether the weights are pre-quantized (like the NF4 HuggingFace builds) and loads accordingly; otherwise loads as bf16. This is the default and usually the right answer.
• bf16 / fp16 / fp32 - force a specific precision for full-precision weights.
• nf4 (quantize) - quantize on the fly with bitsandbytes while loading. Handy if you only downloaded the full BF16 weights and want the 24GB path without re-downloading, though the pre-quantized NF4 files are the smoother route.

The inputs that matter

• model_name - dropdown of subfolders in models/emu35. If it shows "No folders found in models/emu35", that's not a bug - you haven't downloaded the weights yet. That string is the pack's honest way of telling you to go get them.
• vq_model_name - dropdown for the vision tokenizer folder (defaults to vision_tokenizer).
• precision - the five options above.

Outputs

Three wires, and all of them are required by every downstream node:

• model (EMU_MODEL) - the loaded language model.
• tokenizer (EMU_TOKENIZER) - text tokenizer.
• vq_model (EMU_VQ) - the vision tokenizer (the IBQ VQ-VAE with its 262k codebook that turns visual tokens back into pixels).

Install

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/EricRollei/Emu35-Comfyui-Nodes.git emu35
cd emu35
pip install -r requirements.txt

(ComfyUI Manager: search "Emu3.5 Nodes".) Then the weights, which are the real project here:

huggingface-cli download BAAI/Emu3.5-Image --local-dir models/emu35/Emu3.5-Image
huggingface-cli download BAAI/Emu3.5-VisionTokenizer --local-dir models/emu35/vision_tokenizer

Or the NF4 build (wikeeyang/Emu35-Image-NF4) if you're on a 24GB card instead of 48GB+.

The one thing to know

This is the V1 loader and it keeps it simple: no device selection, everything goes to ComfyUI's current torch device. The newer Emu 3.5 Loader V2 adds device, a separate vq_device (so you can park the vision tokenizer on CPU - a known 24GB-VRAM lifesaver), multi-GPU auto placement, and defaults to eager attention, which sidesteps the SDPA-on-Blackwell garbage-output bug. If you're starting fresh, reach for V2. Use this one when you're following an old workflow or you want the least moving parts.