(Down)Load Allegro Model

Every ComfyUI workflow for rhymes-ai/Allegro starts here, because this node is the one that makes the model exist on your disk. The "(Down)Load" in the name isn't marketing: if you haven't grabbed the weights yet, running this node pulls the whole rhymes-ai/Allegro repo down from Hugging Face and then loads it. It's the pack's foundation, and for most people it's also the longest wait of the whole pipeline.

Allegro is a text-to-video diffusion model from rhymes.ai, released in October 2024 - one of the first open "commercial-level" video models, before Wan 2.1 landed in early 2025 and made everything else look dated. It's a 2.8B-parameter diffusion transformer paired with a 175M 3D causal VAE and a big T5 text encoder, and it generates 88 frames (6 seconds at 15 FPS) at 720×1280. If you've followed the video-model scene at all, the honest framing is that Allegro is now a nostalgia pick: Wan does the same job better. But it's Apache-2.0, it runs on an 8GB card with offloading, and this pack is the cleanest way to try it.

How it works

The node reads a directory layout - transformer/, vae/, text_encoder/, tokenizer/ - and loads each piece with the right precision: the VAE goes to fp32, the transformer and T5 encoder to bf16. It wires them into a diffusers-style AllegroPipeline and hands you two things:

• pipe (AllegroPIPE) - the assembled pipeline. This is what the text encoder, both samplers, and the TI2V encoder all consume.
• vae (VAE) - the same VAE, passed out separately because AllegroEncoder and AllegroDecoder take it as their own input.

The path logic is the quirk worth knowing. model_path defaults to models/, which resolves relative to the pack's own folder - custom_nodes/ComfyUI-Allegro/models/ - not ComfyUI's usual models/ directory. If that folder doesn't exist, the node silently kicks off snapshot_download("rhymes-ai/Allegro") and waits for the multi-gigabyte download to finish. That's the "Down" part, and it can look like the UI hung. It hasn't; give it bandwidth and patience, or pre-download the repo yourself and point model_path at it.

The inputs that matter

Of the five string inputs, you will almost always touch exactly one:

• model_path - set this to where the Allegro repo lives. Leave transformer_path, vae_path, text_encoder_path, and tokenizer_path blank unless you're deliberately swapping in a different T5 or VAE; the README is explicit that only the model path should be set normally. Each path must be a folder, not a file - this is from_pretrained territory, so it wants config.json and diffusion_pytorch_model.safetensors inside each directory.

Install

Grab the pack, then let the first run do the heavy lifting:

cd ComfyUI/custom_nodes
git clone https://github.com/bombax-xiaoice/ComfyUI-Allegro
cd ComfyUI-Allegro
pip install -r requirements.txt

Restart ComfyUI. To skip the in-node download, grab the weights separately (git-lfs required):

git lfs clone https://huggingface.co/rhymes-ai/Allegro models

Common issues

The requirements.txt pins hard versions - torch==2.4.1, diffusers==0.28.0, transformers==4.40.1, xformers==0.0.28.post1 - and if you install them over a working ComfyUI environment you can break your existing image nodes. On a dedicated ComfyUI install it's usually fine; on a heavily-modified one, consider only installing the new extras (decord, sentencepiece, imageio, ftfy, bs4, einops) and leaving torch/diffusers alone. And remember: first run isn't hung, it's downloading. Watch for the model folder to grow.