EasyLoader (Mochi)

This is the Easy-Use wrapper for Mochi, Genmo's open text-to-video model. It rolls the checkpoint, the VAE, your prompts, and the video latent setup into one node that outputs a pipe you feed into an easy presampler and sampler - so you skip the multi-node graph that raw Mochi otherwise needs.

Level-setting on Mochi, because it's not a household name: it's a large (~10B-parameter) DiT-based text-to-video model, Apache-2.0 licensed, that Genmo dropped in late 2024. Text in, a few seconds of 848×480 video out. It made a splash as one of the first genuinely open, permissively-licensed video models, but let's be honest about where it sits in 2026 - it's heavy, it's slow on consumer cards, and community attention has drifted hard toward Wan and LTX. If you're specifically here for Mochi, this loader is the clean way to run it. If you're just shopping for a video model, know going in that you're picking one that's more "respectable and open" than "what everyone's actually using this month."

How it works

Mochi generates the whole clip as a batch of video latents conditioned on your text prompt, then decodes them through its own dedicated VAE (this is important - Mochi's VAE is model-specific, not the standard SD one). easy mochiLoader loads the checkpoint and that VAE, encodes your positive/negative prompts, and sizes an empty video latent from your resolution and frame-count settings. The bundled pipe output carries all of that into the sampling stage.

The inputs and outputs that matter

• ckpt_name - the Mochi checkpoint. Required, and the dropdown is empty until you've downloaded one.
• vae_name - defaults to mochi_vae.safetensors. Mochi needs its own VAE; don't point this at a generic SD VAE.
• positive / negative - your text prompts. This is a text-to-video model, so the positive prompt is doing the heavy lifting.
• empty_latent_width / empty_latent_height - default 848 × 480, which is Mochi's native resolution. Straying far from it tends to hurt quality, so change these with care.
• length (default 25, steps of 6) - the number of video frames.

There are also resolution presets, batch_size, and optional model_override / clip_override / vae_override inputs if you want to supply your own loaded components (handy for GGUF or quantized Mochi builds). Outputs: pipe (into easy preSampling → easy kSampler), plus model and vae.

How to install it

Part of ComfyUI-Easy-Use. ComfyUI Manager: search ComfyUI-Easy-Use, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then install.bat (Windows) or pip install -r ComfyUI-Easy-Use/requirements.txt, and restart. The node is light; Mochi is not. You have to download the Mochi checkpoint and its VAE yourself into ComfyUI/models/checkpoints/ and ComfyUI/models/vae/ - the pack ships no models.

Common issues & troubleshooting

Empty ckpt_name. No Mochi checkpoint installed. Download one (the ComfyUI-repackaged Mochi weights are the least painful) and drop it in your checkpoints folder, then restart.

Wrong or missing VAE. If output is garbled color mush, you're probably decoding with the wrong VAE. Mochi requires mochi_vae.safetensors - make sure it's present and selected, not a stock SD VAE.

Out of memory / crawling slowness. Mochi is genuinely big and this is the number-one complaint people hit. Use a GGUF or fp8 build via the *_override inputs, keep the resolution at native 848×480, and start with a short length. If you're on a small card, this may simply be more model than your hardware wants to run.

Off-resolution artifacts. Push width/height far from 848×480 and quality degrades. Stay at or near native and upscale afterward instead.