UNO Model Loader @REDAIGC

REDUNOModelLoader is where the whole ComfyUI-RED-UNO pack earns its reputation as a hassle. Understand that one thing and you're 80% of the way to using it. ByteDance's UNO is an in-context generation model - you hand it a reference image and a prompt, and it puts your object, logo, or character into the scene you describe, no LoRA training, no IP-Adapter. The catch: UNO doesn't speak ComfyUI's model format, so this pack (by QijiTec, the "REDAIGC" Civitai creator) wraps the official Diffusers pipeline in a node. This loader is that wrapper. It's not a load-a-checkpoint-and-go node; it's a Diffusers pipeline wearing a ComfyUI costume, and the costume is the only part that fits natively.

What it does

You get one output: uno_model, a UNO_MODEL object that only its sibling REDUNOGenerate can consume. The inputs are the five things the pipeline needs:

• flux_model - a dropdown of what's in models/diffusion_models. This is where the famous "Missing keys: 236 / Unexpected keys: 236" error lives. You must select the RED-UNO FT FP8 checkpoint from Civitai (model 958009), not a renamed FLUX.1-dev. The README is blunt about it: the model structure is different, so renaming any old FLUX file to fake it will fail.
• ae_model - a dropdown of VAEs in models/vae. It has to be the Diffusers-format VAE (GuangyuanSD/16C_vae_Diffusers on Hugging Face). A normal FLUX VAE will error, because the VAE is decoded inside the Diffusers pipeline, not by ComfyUI's own machinery.
• use_fp8 and offload - tick both. The README claims they "default checked," but the shipped code defaults them to off. On anything under 24GB you want them on; with fp8 + offload the whole thing runs in about 16GB VRAM, and the author reports no visible quality loss versus bf16.
• lora_model - the UNO Dit-LoRA from bytedance-research/UNO, placed in models/loras. Default is None; if you're getting identity-confusion between subjects, loading it is what the FT model was trained to pair with.

Mechanically, on first load it downloads Diffusers-format CLIP and T5 text encoders (around 10GB) into a .cache folder, then builds the pipeline and loads the base model and LoRA (rank 512) off disk. First load is slow; that's the download, not a hang.

Install

Search "ComfyUI-RED-UNO" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/QijiTec/ComfyUI-RED-UNO

Restart ComfyUI. requirements.txt is small - einops, transformers, huggingface-hub, diffusers - Manager handles it, or pip install -r requirements.txt. Then gather the three model files:

• RED-UNO FT FP8 → models/diffusion_models (Civitai: civitai.com/models/958009)
• UNO Dit-LoRA → models/loras (HF: bytedance-research/UNO)
• Diffusers-format VAE → models/vae (HF: GuangyuanSD/16C_vae_Diffusers)

Plus the ~10GB T5/CLIP auto-download on first run.

Gotchas

• "Missing keys: 236" in your console = wrong base model. It means the checkpoint doesn't match UNO's architecture. Swap in the RED-UNO FT file - nothing else fixes it.
• The 10GB cache. The README's own advice: give your ComfyUI virtual environment a dedicated cache dir. On a system-wide ComfyUI install it dumps into your user profile's .cache and eats real estate you didn't expect.
• The VAE must be the Diffusers one. Any FLUX VAE that loads fine everywhere else can still explode here, and the error message won't tell you it's a format problem.

Wire uno_model into REDUNOGenerate.uno_model and you're set. Most of what people report as "this pack is broken" is actually this node fed the wrong checkpoint - get the base right and the rest of the graph behaves.