LLaDA-Image AIO Loader (T8)

LLaDA-Image (the image model from inclusionAI) does not exist in ComfyUI Core yet. No built-in loader knows what its weights look like, so you can't just drag the original release into a CheckpointLoaderSimple. That's the hole this node fills - and it does it in a way that makes your life easy: it treats the whole model as one checkpoint file, text encoder, VAE, tokenizer, and all.

If you came from Stable Diffusion, forget everything about hunting down a separate CLIP or VAE. The "AIO" in the display name is the whole point. Everything LLaDA-Image needs - the diffusion transformer, an LLM-family text encoder, the SigVQ image tokenizer, and the tokenizer bytes - lives inside a single ~49.26 GB BF16 safetensors. You download one file, drop it in models/checkpoints, and pick it from a dropdown. That's the entire setup.

What you're actually loading

The three outputs use ComfyUI's standard port names, but don't let them fool you:

• MODEL - a flow-matching DiT. Not an SDXL-style epsilon predictor, so it doesn't want Karras schedules or CFG 7 (see the scheduler article for the how).
• CLIP - the output type is ComfyUI's generic "text encoder" slot, but this is an LLM-family encoder, not the old CLIP from SD. It's also the host for the SigVQ codebook and the LLaDA2 model that the VQ and edit nodes call on. That's why T8LLaDAImageVQConditioning and T8LLaDAImageEditConditioning connect to this CLIP output instead of a normal text encoder.
• VAE - mostly for VAEDecode at the end, but in the editing workflow it also does the source-image encode, so it feeds the conditioning node too.

There are no other widgets. The single ckpt_name input lists every .safetensors in your models/checkpoints folder, which is exactly where the README tells you to put the model file. Everything else in the workflow branches off these three outputs.

Install

The pack's Manager display name is LLaDA-Image T8 (publisher t8star). Manager install works; note that the registry release 0.1.0 supports only the BF16 files, so if you want the experimental INT8 models, install from GitHub and keep it updated:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-LLaDa-Image-T8.git
cd Comfyui-LLaDa-Image-T8
python -m pip install -r requirements.txt

Then restart ComfyUI and refresh the browser. Requirements are light - tokenizers, tqdm, typing_extensions. No diffusers, no transformers pipeline at inference, and nothing phones home. You do need ComfyUI 0.34.0+ and Python 3.10+.

Grab a model from t8star/LLaDa-Image-Comfy on Hugging Face:

hf download t8star/LLaDa-Image-Comfy LLaDA-Image-Turbo-BF16-AIO.safetensors --local-dir models/checkpoints

Base (50 steps, CFG 5) and Turbo (4 steps, CFG 1) are separate ~49 GB files. You want the one matching the workflow you're running. There are also experimental INT8 versions at ~27.65 GB, clearly labeled, for when disk is the bottleneck.

Where people get burned

• Don't swap in a regular CheckpointLoaderSimple. The README warns about it explicitly; Core has no idea what these tensors are, and you'll get a config-detection error or garbage.
• INT8 files need the current GitHub loader, not the registry build. The old 0.1.0 release predates INT8 support. If a quantized file fails to load, git pull --ff-only inside the node directory and restart.
• One install, not two. The directory has been renamed around a bit - Comfyui-LLaDa-Image-T8 vs the old llada_image_t8 alias. Keeping both copies installed is a classic way to end up with the stale loader winning. Pick one name and stop.
• It's a big model and it's honest about it. Load, VRAM, and offload between GPU and RAM take a while; the developer validated on a 24 GB RTX 5090 laptop with 64 GB RAM, and explicitly does not promise that's the floor. INT8 cuts disk ~44% but is not guaranteed to shrink VRAM or speed up render.

One more thing worth knowing before your first prompt: this is an LLM-encoded model, so write natural sentences, not booru-tag soup. Quality-tag spam does nothing useful here.