RunningHub HeliosModelLoader

Every Helios workflow in this pack starts here. This node loads the entire model - a 14B video transformer, a Wan-style VAE, the scheduler, and the text encoder - from ComfyUI/models/Helios-Distilled/ and hands you a HELIOS_PIPE that feeds the T2V, I2V, and V2V nodes. That's all it does, and it matters more than any other node in the pack, because Helios is heavy and the loader is where you either get it running in ~6GB of VRAM or watch it OOM.

Quick reality check, because the hype train left the station: Helios (from PKU-YuanGroup, March 2026) is the "14B real-time video" model, and "real-time" means 19.5 FPS on a single H100. On your 3090 or 4090 you are not getting real-time anything - but with the loader's offloading modes you can still render minute-scale clips, just not live. That's the honest framing for everything below.

How it actually loads

The node isn't a thin ComfyUI checkpoint loader. It vendors the whole Helios repo and runs the diffusers-style pipeline in-process, so the files don't go in your checkpoint folder - they have to sit in models/Helios-Distilled/ with the exact subfolder layout from the README (transformer/, vae/, scheduler/, tokenizer/, text_encoder/, model_index.json). There's no dropdown to pick a variant: whatever is on disk is what loads. Swap files to switch between Distilled, Base, and Mid.

Under the hood it does a few things worth knowing:

• The VAE is always loaded in fp32, the transformer in your chosen dtype.
• It swaps in flash norms and fp32 RMSNorm, then tries attention backends in order - _flash_3_hub, flash_hub, flash, sage, xformers, native - and uses the first one that actually loads. You don't pick; it probes.
• With low VRAM mode on, it enables group offloading: transformer blocks live on CPU and stream to the GPU as needed. block_level offloads whole blocks, leaf_level goes down to individual layers - leaf uses less VRAM but more CPU bandwidth.
• It caches the loaded pipeline and reuses it across runs as long as your three settings don't change. Change one and it frees the old model, garbage-collects, and reloads. That's why your first run takes minutes and subsequent runs feel snappier.

The three inputs

• weight_dtype - bf16 (default) or fp16. Keep bf16 on anything RTX 30-series or newer. fp16 doesn't save VRAM, just bandwidth, and you'll rarely want it.
• enable_low_vram_mode - on by default, and you should probably leave it on unless you have 24GB+. It's the difference between "runs at all" and "CUDA out of memory."
• group_offloading_type - block_level (default) or leaf_level. The README's rule of thumb: ≤8GB VRAM, use leaf_level at ~6GB; 8–16GB, block_level; ≥24GB, turn offloading off for full speed.

Install

ComfyUI Manager finds it as "ComfyUI_RH_Helios," or:

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

It needs PyTorch ≥ 2.7.1 with CUDA, and the requirements list is long - diffusers, transformers, accelerate, deepspeed, peft, triton, plus a numpy<2.0.0 pin. Here's the trap: the repo's install.sh uninstalls xformers, torchao, and tensorflow, then pins triton==3.3.1 and wandb==0.23.0. Running that script on a ComfyUI install you depend on can break other nodes. Do the manual pip install -r requirements.txt in a venv instead.

Then grab the model (~tens of GB - the community first-wave reports were talking 80GB for an unofficial quant, the real bf16 is bigger):

pip install "huggingface_hub[cli]"
huggingface-cli download BestWishYsh/Helios-Distilled --local-dir ComfyUI/models/Helios-Distilled

Use Helios-Distilled, the x0-prediction checkpoint the pack is tuned around. Helios-Base is the quality pick but uses v-prediction with standard CFG, and Helios-Mid is, by the README's own admission, not worth your disk.

Where people get burned

The models/Helios-Distilled path is hardcoded - if your model_index.json is missing or a subfolder is wrong, you get a FileNotFoundError that doesn't say which piece is missing. And if you're on an 8GB card, remember that "~6GB with group offloading" assumes the offloading is on and working; the first load can still spike before the offload settles. Give it a minute - the progress bar doesn't lie about loading, but ComfyUI will look frozen while the 14B model is being read off disk.