ADMD_LoadLora

Everything else in this pack is about making a motion LoRA. ADMD_LoadLora is about using it. It takes the path of a trained temporal LoRA (the kind ADMD_SaveLora produces), wraps it with a strength, and hands you a MOTION_LORA that plugs straight into AnimateDiff-Evolved's motion-LoRA loader - the standard AnimateDiff inference path. Train in this pack, generate in AnimateDiff-Evolved; the README says that's exactly what the output is for.

The input is a plain lora_path string - the relative path under ComfyUI/models/animatediff_motion_lora/. The file you'd paste here looks like 2026-08-16/14-30-22/my_lora/500_my_lora_r64_temporal_unet.safetensors, which is precisely what ADMD_SaveLora returns in its lora_path output. Note it's forceInput, so you're expected to wire it rather than type it - hook the save's STRING output straight in and there's nothing to transcribe.

Inputs and output

• lora_path - relative path under models/animatediff_motion_lora/. Only the temporal LoRA goes here; that's the file AnimateDiff actually consumes. If the path doesn't resolve, the node raises a clear FileNotFoundError, which beats a silent no-op.
• strength (default 1.0, range 0–10) - how hard the motion applies. This is a motion LoRA, not a style LoRA: 1.0 is the honest starting point, and you tune down (0.4–0.7) when the learned motion fights your prompt's intended movement. A motion LoRA at full strength can overpower the checkpoint's own behavior.
• prev_motion_lora (optional) - a previous MOTION_LORA. Feed one in and this node appends to it, letting you stack multiple trained motions in one loader. The underlying list is cloned, so earlier nodes in the chain aren't mutated.

Output: a single MOTION_LORA, which goes into AnimateDiff-Evolved's motion LoRA loader (the AD_LoadMotionLora-family node) alongside your regular AnimateDiff motion module. That's the whole point of the use_motion_lora_format save on the training side - it was written so this handoff works.

Install and gotchas

Same pack install as its siblings (Manager → "ADMotionDirector", or clone + pip install -r requirements.txt into the pack dir). The one hard dependency for this node's output to be useful is ComfyUI-AnimateDiff-Evolved - without it there's no MOTION_LORA consumer. ADMD_LoadLora itself is a featherweight: it constructs a small MotionLoraInfo(name, strength) record and appends it to a list; no model is loaded here, so it costs nothing to sit in the graph.

Two honest notes. First, this node does not verify the LoRA is valid for your checkpoint - motion LoRAs are architecture-locked to SD 1.5's UNet, so if your inference checkpoint isn't SD 1.5 you'll get garbage or an error downstream, not here. Second, the wider reality check: AnimateDiff is a superseded 2023-era ecosystem in 2026 - this path is for people deliberately chasing the AnimateDiff motion look, and if that's you, this is the cleanest handoff in the pack.