(Down)Load Allegro TextImage2Video Model

The text-image-to-video equivalent of LoadAllegroModel, with one genuinely useful twist: it doesn't have to cost you a second full model download. LoadAllegroTI2VModel assembles the TI2V pipeline - the transformer that powers Allegro's image-to-video mode - but if you already have the plain Allegro weights, most of the heavy pieces are shared. This is the "(Down)Load" node for the Allegro-TI2V repo, and the README is unusually thoughtful about disk space for a change.

How it works

Same shape as its T2V sibling: you give it a model_path (default ti2v_models/, again resolved relative to the pack's own folder), and if that directory doesn't exist it downloads rhymes-ai/Allegro-TI2V from Hugging Face on first run. It loads the AllegroTI2V transformer (the 2.8B DiT variant trained for reference-frame conditioning), plus the shared VAE, T5 tokenizer, and T5 encoder - VAE in fp32, transformer and text encoder in bf16 - and returns a pipe (AllegroPIPE) plus the vae (VAE) as separate outputs.

Here's the good part. The TI2V transformer is genuinely new, but the VAE, text encoder, tokenizer, and scheduler are the same components Allegro's text-to-video mode already uses. So instead of downloading the full Allegro-TI2V repo, the README shows you how to grab only the new transformer and symlink the rest to your existing models/ folder:

mkdir -p ti2v_models/transformer/
wget https://huggingface.co/rhymes-ai/Allegro-TI2V/resolve/main/transformer/config.json \
     -O ti2v_models/transformer/config.json
wget https://huggingface.co/rhymes-ai/Allegro-TI2V/resolve/main/transformer/diffusion_pytorch_model.safetensors \
     -O ti2v_models/transformer/diffusion_pytorch_model.safetensors
ln -s ../models/vae ti2v_models/vae
ln -s ../models/text_encoder ti2v_models/text_encoder
ln -s ../models/tokenizer ti2v_models/tokenizer
ln -s ../models/scheduler ti2v_models/scheduler

That gets you TI2V for the price of one ~5GB transformer instead of the whole repo. The symlink paths above are the README's shape - adjust to wherever your pack folder actually lives.

The inputs that matter

• model_path - the TI2V model directory. The only field you'll normally set.
• transformer_path / vae_path / text_encoder_path / tokenizer_path - leave blank unless you're swapping components; if the node finds the standard layout under model_path it wires them all up itself.

It returns pipe (→ text encoder, TI2V encoder, TI2V sampler) and vae (→ AllegroDecoder/Encoder).

Install

The pack install is the same as the T2V side:

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

Then let the first run download rhymes-ai/Allegro-TI2V, or pre-fetch it with git lfs clone, or use the symlink trick above if you already have Allegro.

Common issues

Two things to keep straight. First, TI2V mode wants its own loader - don't feed the TI2V nodes a pipe from LoadAllegroModel; the transformer type differs (the T2V pipe lacks the reference-frame conditioning path) and the example workflows use the matching loader for a reason. Second, watch the pinned requirements.txt (torch 2.4.1, diffusers 0.28.0, transformers 4.40.1) - on a heavily customized ComfyUI, installing it wholesale can clash with existing image nodes. If the rest of your stack is already working, install just the new extras and leave the big pins alone.