Load EasyAnimate Model

Every EasyAnimate workflow in this pack starts here, and it's the node doing the most work. While the samplers are the stars, LoadEasyAnimateModel is the thing that builds the entire pipeline - VAE, the 3D transformer DiT, the text encoder(s), the scheduler - and hands it to a sampler as a single easyanimate_model bundle. It's also the node where your whole night gets won or lost, because every option on it is a memory or quality tradeoff.

EasyAnimate is Alibaba's PAI team's open-weights video model family (Apache 2.0), a sibling to Wan and Qwen in the Chinese open-weights wave. The model itself is a DiT-style transformer with a temporal VAE, and it's gone from V1 to V5.1: the current V5.1 line uses a Qwen2-VL text encoder and flow-matching sampling, 49 frames at 8fps. This node has to know all of that, which is why it's a form, not a switch.

The inputs that actually matter

model - a 17-option dropdown, default EasyAnimateV5.1-12b-zh-InP. Read it like: version, size, then role. -InP is the image/video-to-video inpaint weights, -Control is the control-model weights (Canny, pose, depth, trajectory, camera), and the plain -zh is text-to-video. You can also pick older V3/V4/V5 checkpoints here, so the whole history of the pack is one dropdown. Start with the default.

GPU_memory_mode - the three-way choice that determines whether this runs at all:

• model_cpu_offload (default) - the model shuffles between GPU and CPU; the fastest of the three on consumer cards.
• model_cpu_offload_and_qfloat8 - same offloading but quantizes the transformer to float8. Big VRAM savings, small quality cost. This is the one people actually land on.
• sequential_cpu_offload - layer-by-layer offload. Slowest by a mile (community reports in the ~15s/it range on 24GB cards) but it fits almost anything.

The 12B is genuinely VRAM-hungry. On a 24GB card the community consensus is: if model_cpu_offload OOMs, flip to qfloat8 before you give up. The 7B models are far friendlier and the sane choice on 16GB.

model_type - Inpaint or Control. Inpaint is the default and serves all the text/image/video samplers. Switch to Control only when you've loaded a -Control checkpoint. They build different pipelines (the control pipeline additionally consumes control videos, reference images, and camera conditions), so matching this to your weights matters more than it looks like it does.

config - four YAML files, one per model generation (v3, v4, v5, v5.1). The default easyanimate_video_v5.1_magvit_qwen.yaml matches the default model. If you're using an older checkpoint, pick its config - this is what tells the node which VAE, text encoder and transformer structure to construct.

precision - bf16 or fp16. bf16 is the default and right on modern cards. If your GPU doesn't support bfloat16 - think 2080 Ti, V100 - switch to fp16 or the model won't behave (the README calls this out explicitly).

What comes out

One output: easyanimate_model, typed EASYANIMATESMODEL. It's an opaque bundle containing the loaded diffusers pipeline, dtype, model name, and an empty LoRA list that LoadEasyAnimateLora can append to. Wire it into any of the six samplers in the pack.

Installing and making it work

Install via ComfyUI Manager (search "EasyAnimate") or clone the repo to ComfyUI/custom_nodes/EasyAnimate/ and run:

cd ComfyUI/custom_nodes/
git clone https://github.com/aigc-apps/EasyAnimate.git
cd EasyAnimate/
pip install -r comfyui/requirements.txt

The model files are the real install. Drop each checkpoint into ComfyUI/models/EasyAnimate/<model_name>/ - the V5.1 7B line is ~30GB, the 12B line ~39GB, and the README wants about 60GB of free disk to be comfortable. The node reads folder_paths.models_dir/EasyAnimate first, with a fallback to a cache dir. Dependencies are heavy: it pulls diffusers and transformers into your shared ComfyUI Python environment, so watch for version conflicts with other video packs.

Common pain points, all real: OOM at load on 24GB cards with the 12B (see the memory modes above); a big system-RAM spike while weights load (people report tens of GB - make sure your pagefile/swap can take it); and loading a -Control checkpoint while leaving model_type on Inpaint, which gives you confusing downstream failures. None of these are the node misbehaving - they're the price of a 12B model on a consumer box. If you just want to make a video tonight, the 7B -InP with model_cpu_offload_and_qfloat8 is the setup that gets you there.