UNO Model Loader

Don't expect a normal checkpoint loader. UNO Model Loader is the front door to ByteDance's UNO, an in-context generation model that rides on top of FLUX.1-dev, and it drags in five pieces of a model in one go. Wire it up, and the payoff is subject-preserving generation - keep a character or a product consistent across scenes without training a LoRA - which is the Flux-era answer to what IP-Adapter did for SD 1.5 and SDXL (IP-Adapter never made it onto Flux). It slots next to PuLID for faces and Kontext for instruction editing, but UNO's trick is different: it learned to read reference images straight through the DiT's context window.

How it works

The node is a thin ComfyUI wrapper around the official bytedance/UNO pipeline. It assembles a FLUX.1-dev DiT, the AE autoencoder, both text encoders (T5 for the long prompt path, CLIP for the pooled guidance vector), and - the load-bearing bit - the UNO LoRA. That LoRA, dit_lora.safetensors from bytedance-research/UNO, is what turns plain Flux into UNO: it injects the in-context reference-image conditioning into the model's weights (rank 512, set at load time). Leave lora_model on "None" and you've basically just loaded vanilla Flux dev, so in practice you always want it selected. If it's missing, the pack's own code tries to auto-fetch it from Hugging Face, so a bare dit_lora path often just works.

The inputs that matter

• flux_model - the FLUX.1-dev checkpoint, picked from models/unet. You can run the full dev weights or the fp8 variant; the loader builds a flux-dev or flux-dev-fp8 pipeline accordingly.
• ae_model - ae.safetensors from models/vae.
• t5_model and clip_model - both pull from models/clip, which is the easiest thing to fumble. The author added code that detects swapped T5/CLIP files and swaps them back for you, which tells you exactly how common that mix-up is.
• use_fp8 (default off) - loads the fp8 Flux variant. With offload (default off) - which keeps the text encoders and AE on CPU and shuttles them to the GPU as needed - the README's claim is 24GB VRAM. On 12GB it's rough; that's the community verdict, not marketing.
• lora_model - the UNO LoRA, from models/loras. Keep it set.

The output is a single uno_model of type UNO_MODEL, and its only consumer in the pack is UNO Generate.

Installing it

ComfyUI Manager: search "ComfyUI UNO Nodes". Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/alexgenovese/ComfyUI-UNO-Flux

Restart ComfyUI, then drop the models where the loader looks for them. requirements.txt asks for einops, transformers, huggingface-hub, and diffusers - ComfyUI usually already has most of these, so a failed first run is often just a missing pip install -r requirements.txt.

flux1-dev.safetensors      → models/unet
ae.safetensors             → models/vae
t5xxl_fp16.safetensors     → models/clip   (xlabs-ai/xflux_text_encoders)
clip_l.safetensors         → models/clip
dit_lora.safetensors       → models/loras  (bytedance-research/UNO)

The README's "clip and t5 will autodownload" is half-true: the base text encoders (openai/clip-vit-large-patch14 and a T5) get pulled from Hugging Face on first load. It's a chunky first run, and you still need the xlabs T5 file in models/clip for the actual UNO T5 weights.

Where people get burned

Beyond the T5/CLIP mix-up: an empty models/unet or models/vae means the dropdowns are empty, so nothing errors until you load a workflow and the widget values point at files you don't have. And the licensing is a genuine constraint - UNO's weights are CC BY-NC 4.0 on top of FLUX.1-dev's non-commercial terms, so this is a research-toy-or-hobbyist tool, not something to build a paid product on.