Load EasyAnimate Lora

LoadEasyAnimateLora is the pack's answer to one question: how do I get my trained or downloaded LoRA onto an already-loaded EasyAnimate model without reloading 30GB of weights? The answer is clever - it doesn't merge anything at load time at all.

You feed it the easyanimate_model bundle from LoadEasyAnimateModel, pick a LoRA, set a strength, and it hands the same bundle back with the LoRA's path and strength appended to a list. The actual weight merge happens inside the sampler, right before sampling, and gets un-merged right after. That means the loaded model in memory stays pristine and you can chain two, three, four of these nodes in a row - each one just appends to the pile.

The inputs

• easyanimate_model - the bundle from the loader. Non-negotiable.
• lora_name - a dropdown populated from your ComfyUI models/loras folder (the standard safetensors location). EasyAnimate's own EasyAnimateV5-Reward-LoRAs from HuggingFace and anything you trained with the repo's scripts land here the same way.
• strength_model - default 1.0, but the range is wild: -100 to 100 in 0.01 steps. You'll live in the 0.5–1.5 band; negative strengths exist if you want to invert a style.
• lora_cache - the toggle that makes this node interesting.

lora_cache is the feature worth understanding before you build a big workflow. When it's on, the node snapshots the transformer's original weights to CPU RAM the first time. Every subsequent run, switching from one LoRA to another means loading the cached clean weights and merging the new LoRA on top - no repeated weight juggling, much faster iteration if you're A/B testing styles. The cost: it holds a full copy of the transformer in system memory, and that's a chunky chunk of RAM on a 12B model. On a machine with 32GB+ of RAM it's a great trade; on a tight box, keep it off.

With lora_cache off (the default), the sampler merges, runs, then un-merges every single time. Correct and safe, but slower per run, and if you change strengths between runs you're paying the merge cost again.

Where it sits in the workflow

LoadEasyAnimateModel → LoadEasyAnimateLora → EasyAnimateV5_T2VSampler

It must come between the loader and the sampler in the chain - the bundle flows through it. It works with every sampler in the pack, V3 through V5.1, since it just augments the same EASYANIMATESMODEL type.

Gotchas

The merge lives in the sampler, not this node, so if you don't wire it before a sampler the LoRA silently never happens. And because EasyAnimate's samplers un-merge after each run when caching is off, a LoRA-heavy workflow is inherently doing a lot of weight math behind the scenes - that's normal, not a bug, and it's exactly the tax lora_cache exists to avoid. Watch system RAM if you enable the cache on a 12B checkpoint, and remember that LoRA strength stacking still applies - two 1.0s is not 2.0, it's usually mush.