RunningHub USO Loader

The RunningHub USO Loader has no inputs, no settings, and one job: load the entire USO stack into memory and hand it to the sampler. It looks like the least interesting node in the pack - and it kind of is - but it's also where 90% of setup pain lives, because "loading the USO stack" means fetching FLUX.1-dev, the VAE, two text encoders, a SigLIP vision model, and ByteDance's USO adapters into exactly the right folders first.

What it does

USO (arXiv 2508.18966) is ByteDance's unified style-and-subject generation model, released open-source in late August 2025 and ported to ComfyUI by RunningHub - the cloud GPU platform that runs ComfyUI in your browser, hence the "RH_" naming. The model rides on FLUX.1-dev as a LoRA plus a projector, with a SigLIP vision encoder reading your style references. The Loader is the node that assembles all those pieces:

• FLUX.1-dev base transformer (loaded in FP8) and its VAE
• FLUX's T5 and CLIP text encoders
• SigLIP vision model (hooked in as the pipeline's vision encoder)
• The USO dit_lora + projector adapters, hardcoded to rank 128 with CPU offload on

Every one of those settings - flux-dev-fp8 model type, lora_rank=128, offload - is baked into the source. There's nothing to configure, which is why the node takes no required or optional inputs. Its single output, USO Modules (type RHUSOMudules), is a dict bundling the SigLIP processor and the ready pipeline. You wire that into the RunningHub USO Sampler's uso input and you're done.

The real install is the models, not the node

The code never downloads anything (hf_download=False is hardcoded). If a file isn't in the right folder, you get a load error - no auto-fetch. So the honest install sequence is: install the pack, then place the models exactly where the source looks for them:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_USO
cd ComfyUI_RH_USO && pip install -r requirements.txt

Then the model files (all under ComfyUI's models/ dir):

diffusers/FLUX.1-dev/flux1-dev.safetensors   # the FLUX.1-dev base
diffusers/FLUX.1-dev/ae.safetensors          # the VAE
uso/uso_flux_v1.0/dit_lora.safetensors       # USO LoRA
uso/uso_flux_v1.0/projector.safetensors      # USO projector
clip/siglip-so400m-patch14-384/              # SigLIP vision model

Here's the trap: the README's download list stops there, but the source unconditionally loads two more things - the T5 encoder from models/clip/xflux_text_encoders and the CLIP encoder from models/clip_vision/clip-vit-large-patch14. If you already run FLUX in ComfyUI you probably have both (they're standard FLUX parts). If you don't, grab xlabs-ai/xflux_text_encoders and openai/clip-vit-large-patch14 before your first run, or the Loader dies mid-load.

Also worth knowing: requirements.txt is lifted wholesale from the upstream USO repo. It drags in deepspeed, gradio, onnxruntime-gpu and an openai/CLIP build from git that none of the node code ever imports. If pip starts fighting your existing environment over those, the two nodes genuinely only need the smaller core set - torch, transformers, diffusers, accelerate, safetensors, numpy. ComfyUI Manager (search "ComfyUI USO Node") handles the clone; just don't assume every line of that requirements file is load-bearing.

What to expect the first time

First load is slow and chunky - it pulls all four+ models to CPU, builds the FP8 pipeline, then offloads, and it does this every time the queue runs. That's a minute or two of "not responding" that isn't a hang. Budget ~35GB of disk and ~16GB of VRAM at generation time. And if you see an error naming a missing file, it's almost always one of the two text encoders the README forgot to mention.