Walkyrie Loader

Walkyrie Loader is the front door to a whole mini-ecosystem: a ~1.3B text-to-image model called Walkyrie, wrapped for ComfyUI by its own trainer. It's a real underdog story. The author took Wan 2.1's lightweight 1.3B video model, pruned the giant UMT5 text encoder down to about 1B parameters, and retrained the whole thing for still images. Then they shipped the result as a single ~2.8 GB merged checkpoint and built this three-node pack around it. Loader loads that file; that's the whole job. Nothing here calls an API and nothing needs a key - it's one safetensors file plus some config JSON that ships inside the pack.

Why you'd reach for it: Walkyrie is aimed squarely at people whose GPUs laugh at 14B video models. A 1.3B diffusion transformer with a pruned text encoder will run on 6–8 GB VRAM class hardware, which is the entire pitch. The author said so themselves: they wanted a small model people with weaker cards could actually use. Just set expectations - it's released as a "Preview" trained to roughly 20% of its budget, and the trainer openly admits anatomy is the weak spot. This is an indie experiment with real charm, not a finished frontier model. Treat it as a fun, fast, uncensored-by-default toy that improves the more people test it.

How it works

The Loader expects a merged single-file checkpoint in ComfyUI's models/checkpoints folder - not a diffusers folder, one .safetensors with keys prefixed transformer.*, text_encoder.*, and vae.*. When you pick one, the node splits that state dict into the three components, builds each model from the config files bundled in the pack, and assembles a pipeline_walkyrie around them: a Wan transformer, the Wan VAE, UMT5 text encoder, and a flow-matching Euler scheduler. Your file basically gets unpacked into a standard Wan-architecture pipeline.

One detail worth knowing: the pipeline is cached. Load the same checkpoint at the same dtype again and you get the identical pipe back, no reload. Load a different one and the old pipe is evicted from VRAM first. So switching checkpoints mid-session is cheap; just don't expect two models to coexist in memory.

The inputs that matter

• ckpt_name - dropdown of everything in models/checkpoints. If it's empty, you haven't downloaded the model yet (see below).
• dtype - bfloat16 (default), float16, or float32. Keep the default on anything Ampere or newer; the FP8 checkpoint you'll download gets cast to this precision on load anyway.

The output is a single WALKYRIE_PIPE - a custom pipe type you wire into the Walkyrie LoRA Loader or straight into Walkyrie Sampler. No other node understands it, which is fine, because those are the only two places you need it to go.

Installing it

Through ComfyUI Manager, search for walkyrie, or clone it manually:

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

Then restart ComfyUI. The pack pins diffusers==0.34.0 and transformers==4.46.3 in its requirements.txt, and it imports AutoencoderKLWan and WanTransformer3DModel from those versions, so let Manager install (or downgrade) those without complaint.

The part the README won't tell you (because the README is literally just a screenshot): grab the model from Hugging Face:

# into ComfyUI/models/checkpoints/
wget https://huggingface.co/kpsss34/Walkyrie-1.3B-v1.0/resolve/main/walkyrie_FP8e4m3fn.safetensors

That's the walkyrie_FP8e4m3fn.safetensors file, about 2.8 GB. Drop it in the checkpoints folder, hit refresh, and it'll appear in the ckpt_name dropdown.

Troubleshooting

• Empty ckpt dropdown - the file isn't in models/checkpoints, or ComfyUI didn't rescan. Restart, or check the filename isn't inside a subfolder.
• Dependency conflicts - the diffusers pin can fight other Wan-related nodes (Kijai's wrapper and friends) that want newer versions. If stuff breaks after install, it's usually this; a dedicated venv or accepting the downgrade is the fix. This is the classic ComfyUI custom-node dependency mess, in miniature.
• New-model surprises - since it's a preview, quality varies. Don't judge it on one seed; give the Sampler's guidance_scale a nudge and try a few prompts before writing it off.