Alice T2V Loader (Eric)

This is the boring half of a two-node pack, and that's exactly the point. Alice T2V is Mirage's open-source 14B MoE text-to-video model - a T5 text encoder, two DiT experts, and a VAE - and something has to drag all of that into memory before a single frame gets generated. This loader is that something. It's not the camera; it's the generator's power supply, and it's where 90% of the work (and every setup mistake) actually lives.

The loader's only job is to build an AliceTextToVideo pipeline from the model folder and hand it out as a single pipeline output. That output is typed ALICE_PIPELINE, and the only thing in the wild that accepts it is the pack's other node, Alice T2V Generator (Eric). So the two always travel together: loader in, generator after it, frames out the far end.

How it works

Under the hood it's a module-level cache keyed on all five settings. The pipeline is built once and held in memory, so your second and third generations are fast. But change any setting - flip t5_cpu, pick a different GPU - and it evicts the old pipeline and reloads from scratch, which is a minute-plus of silence on a big model. Don't tickle the flags between runs; set them once and leave them.

It also does two quiet self-healing jobs you'd otherwise be fixing by hand. First, if the umt5-xxl tokenizer isn't in your model folder, it downloads it (~2 MB) into ckpt_dir/google/umt5-xxl. Second, the HuggingFace upload ships sharded safetensors without the diffusion_pytorch_model.safetensors.index.json that diffusers needs to find its weights - the loader reads each shard's header (metadata only, no tensors loaded) and generates that index for you. Nice touch for a wrapper that otherwise expects you to do everything yourself.

The inputs that matter

• ckpt_dir - the only one you'll get wrong. Point it at the folder you downloaded gomirageai/Alice-T2V-14B-MoE into. It must contain low_noise_model/ and high_noise_model/ subfolders, or the loader throws "Expected subfolder not found."
• offload_model (default True) - swaps whichever DiT expert isn't active to CPU during each step. This is the difference between ~28 GB and ~56 GB VRAM. Disable it only on 96 GB+ cards.
• t5_cpu (default True) - keeps the ~6B T5 text encoder on CPU. Encoding is quick, and it frees VRAM for the DiTs. Leave it on.
• convert_model_dtype - casts weights to bfloat16, halving DiT VRAM at a minor quality cost. Only for lower-VRAM GPUs.
• device_id - CUDA index, 0 to 7. You'll only touch this on multi-GPU rigs.

Installing this pack

The README's whole point: this is a wrapper, and the model does not come with it. Three downloads and one pip install, in order:

# 1. The weights (~27 GB) - remember this path, it's your ckpt_dir
pip install huggingface_hub
huggingface-cli download gomirageai/Alice-T2V-14B-MoE --local-dir "D:/models/Alice-T2V-14B-MoE"

# 2. The pack itself
cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric-Alice-T2V-ComfyUI-Wrapper

# 3. The Alice source code has to be "vendored" into the pack (it's not a pip package)
cd Eric-Alice-T2V-ComfyUI-Wrapper
git clone https://github.com/mirage-video/Alice.git /tmp/Alice
python setup_vendor.py --alice-src "/tmp/Alice"

# 4. The only extra Python dependency
pip install easydict

Then restart ComfyUI. ComfyUI Manager can install the repo itself (it's on the registry), but that's a trap: Manager won't run setup_vendor.py, so the nodes will silently fail to load until you do. The vendoring step is mandatory, not optional. Everything else - torch, transformers, diffusers - already ships with ComfyUI.

Common issues

• "Cannot import alice package" - you skipped or botched Step 3. Re-run setup_vendor.py and confirm you see ✓ Vendor install verified - alice imports OK. Point --alice-src at the repo root (the folder containing alice/), not a subfolder.
• "Expected subfolder not found" - wrong ckpt_dir, or a partial download. The full folder has low_noise_model/ and high_noise_model/ plus T5 and VAE shards.
• First load feels frozen - it's the tokenizer download plus shard-index generation plus a 14B pipeline actually loading. That's normal, once.
• VRAM explodes at 70 GB+ - you turned off both offload_model and t5_cpu on a card that can't hold it.

Realistic expectations: this model is brand-new (early 2026), the Reddit footprint is thin, and the few threads there are call it "unknown tier" - a fresh 14B MoE video model without Wan's LoRA or ControlNet ecosystem yet. If you're on a 24 GB card, offload_model=True + t5_cpu=True + convert_model_dtype=True is the config that fits. Load once, stop touching it, and let the generator do the talking.