Load A2 Model

LoadA2Model is where the real weight of ComfyUI-SkyReels-A2 lives, and I mean that literally. While the other nodes in this pack are thin text-and-image pass-throughs, this one loads Skywork's SkyReels-A2 pipeline - a 14-billion-parameter fine-tune of Wan 2.1 - into memory, and it's the first thing to go wrong in almost every "why won't my workflow run" thread about this pack.

Point it at a local copy of the model and it returns the a2_model that A2VideoGenerator feeds on. There's no download built in; you have to fetch the weights yourself first.

How it works

The node expects pipeline_path to point at a HuggingFace-style diffusers directory containing three subfolders: image_encoder/ (a CLIP vision model, loaded in float32), vae/ (the Wan autoencoder, also float32), and transformer/ (the A2 transformer, loaded with use_safetensors=True at your chosen dtype). It then assembles those into Skywork's A2Pipeline and swaps in a UniPCMultistepScheduler configured for flow-matching (flow prediction, flow_shift=8), which is the right sampler setup for a Wan-family model - you don't want to override it.

So despite the MODEL output type, this is not a ComfyUI-native model. It's a diffusers pipeline object. It will not plug into KSampler or any standard ComfyUI node, and mixing it up is a common source of confusion.

Inputs that matter

• pipeline_path - the directory you downloaded the model to. It must be the repo root (where image_encoder/, vae/, transformer/ live), not a subfolder. Get the model with:

  git lfs install
  git clone https://huggingface.co/Skywork/SkyReels-A2 /path/to/model
  

  That's a 14B transformer plus VAE plus encoder - a multi-tens-of-GB download, so have disk space and patience.
• dtype - bfloat16 is the sane choice on any modern GPU; float32 roughly doubles the memory of an already-enormous model for no visible benefit at 480p.
• device - cuda. CPU works in theory; in practice a 14B flow-matching model on CPU is a screensaver.

Installing it

Same as every node in this pack: install Yuan-ManX/ComfyUI-SkyReels-A2 via ComfyUI Manager (search "SkyReels") or

cd ComfyUI/custom_nodes
git clone https://github.com/Yuan-ManX/ComfyUI-SkyReels-A2.git
cd ComfyUI-SkyReels-A2
pip install -r requirements.txt

Restart, then download the model as above. The requirements pin diffusers from git and include flash_attn, so the install is the sort that can eat an afternoon, especially on Windows.

Common issues

The big three: the model repo isn't downloaded (nothing at pipeline_path → a FileNotFoundError from diffusers), wrong dtype for your VRAM (bf16 on a 14B model plus fp32 VAE/encoder realistically wants 24GB-class hardware - people do run it lower with aggressive offloading, but don't expect comfort), and version drift - because the pack rides git-master diffusers, a pip install -U diffusers from something else can silently break the load. The README only lists A2-Wan2.1-14B-Preview as available; the "full" and "Infinity" variants were still marked "to be released", so don't hunt for them.