Walkyrie LoRA Loader

Walkyrie LoRA Loader is the optional middle stop in the Walkyrie pipeline: it takes the WALKYRIE_PIPE from the Loader, stacks a LoRA into it, and hands the same pipe back so the Sampler can run with the LoRA baked in. There are only three nodes in this pack, and this is the only one that changes the model - so if you want a character, a style, or a quality boost out of the base Walkyrie checkpoint, this is your only lever.

Since Walkyrie is a derivative of Wan 2.1's 1.3B video model, Wan-family LoRAs are the natural match. The Wan LoRA scene is huge and most of it is trained on the 1.3B or 14B video variants, so this is where the pack plugs into an ecosystem that already exists - same musubi-tuner and diffusion-pipe trained LoRAs, different harness. Don't expect every Wan LoRA to transfer cleanly to a retrained T2I model, but the architecture is close enough that it's usually worth trying.

How it works

Two calls under the hood, straight out of the diffusers reference usage:

pipe.load_lora_weights(lora_path)
pipe.fuse_lora(lora_scale=lora_scale)

load_lora_weights merges the LoRA into the model's weights in memory, and fuse_lora bakes it in permanently - it's not an inference-time overlay. That's the gotcha worth remembering: the fuse sticks. Because the pipeline is cached by the Loader, once you fuse a LoRA it stays fused for every subsequent sample. Change the prompt, change the seed, re-run the graph - the LoRA is still in there. If you want it out, you have to evict the cache by loading a different checkpoint (or the same one at a different dtype) or restarting ComfyUI. There's no "unfuse" button here, so treat the LoRA Loader as a one-way door for the session.

The inputs that matter

• pipe - the WALKYRIE_PIPE from the Loader. Must be connected first.
• lora_name - dropdown of LoRAs in models/loras. Drop your .safetensors files there and refresh.
• lora_scale - float from 0.0 to 2.0, default 1.0. Lower it toward 0.4–0.7 for subtle style or quality tweaks; keep 1.0 for character/identity LoRAs you trained yourself. Past 1.5 things usually get crunchy.

Output is the same WALKYRIE_PIPE type, one step closer to the Sampler.

Installing it

It ships in the same pack as the Loader, so you already have it:

cd ComfyUI/custom_nodes
git clone https://github.com/kpsss34/walkyrie

or search walkyrie in ComfyUI Manager, then restart. It needs the Walkyrie checkpoint loaded first, and the pack's pinned diffusers==0.34.0 / transformers==4.46.3 still apply - if those got upgraded past what the Wan classes expect, the node breaks before it ever sees your LoRA.

Troubleshooting

• LoRA dropdown is empty - files aren't in models/loras or ComfyUI didn't rescan after you dropped them in.
• "Can't seem to swap" a LoRA mid-session - that's the cache, not you. Change checkpoint or dtype to evict the pipe, or restart.
• LoRA barely does anything or looks broken - try a different scale, and remember this is a preview model trained on a fraction of its budget; a LoRA trained for a different Wan variant is a coin flip.