Walkyrie Sampler

This is the payoff node. The Loader loads the model, the LoRA loader optionally reshapes it, and Walkyrie Sampler is where the actual image gets made. You feed it the WALKYRIE_PIPE, type a prompt, hit go, and out comes a ComfyUI IMAGE tensor - a normal still image you can wire into Save Image, Preview, upscalers, whatever your usual downstream looks like.

Worth being clear about the "video" elephant in the room: Walkyrie descends from Wan 2.1's video architecture, but the author retrained it for images, and this node only ever produces one frame. The latent is hard-coded to a single frame in the pipeline; there's no frame-count input and no video output. Wan video models making great stills was a known community discovery, and Walkyrie is essentially that trick made permanent and shrunk down to fit small GPUs. If you came here expecting clips, that's not this pack.

How it works

The Sampler calls the pipeline's __call__ directly with your prompt, dimensions, steps, guidance, and seed, running the standard flow-matching denoise loop with a Wan transformer, then decodes the latents through the Wan VAE into an image. It also hooks ComfyUI's progress bar so you get a real progress readout instead of a frozen UI. The prompt is processed through the pruned UMT5 text encoder - max ~226 tokens, so long prose gets truncated; be concise and descriptive rather than rambling.

One thing the defaults hide: guidance_scale defaults to 1.0, which switches classifier-free guidance off entirely. In the pipeline, CFG only kicks in when the value is above 1.0, and at 1.0 it skips the negative-prompt pass for speed. That's why default output can feel flatter and less prompt-loyal than you expect. The Wan community consensus is that this architecture wants CFG around 3.5–5 for real prompt adherence and detail, so bump it up - 4–5 is a sane starting point. You're giving up a bit of speed, but you're getting a model that actually listens.

The inputs that matter

• prompt - multiline, so write naturally. Descriptive captioning style works best.
• width / height - 64–4096, step 16, default 1024×1024. Must stay divisible by 16 (the step size enforces that in the UI, and the pipeline hard-fails otherwise). Wan T2I will do 1080p+ if your VRAM allows; on a small card, 768×768 or 832×832 is a smarter default than the full 1024.
• steps - default 20. Fine for this model; 20–30 is the quality sweet spot. More steps buys diminishing returns fast.
• guidance_scale - the one that matters most, see above. 0–20, step 0.1.
• seed - default 0; set one to reproduce a result. Standard.

Output is a single IMAGE tensor.

Installing it

Same pack as the other two, so you're already there:

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

or install walkyrie from ComfyUI Manager, then restart. It needs the checkpoint in models/checkpoints (walkyrie_FP8e4m3fn.safetensors, ~2.8 GB from Hugging Face) before the Loader has anything to pick.

Troubleshooting

• Output looks weak / ignores the prompt - you're almost certainly on CFG 1.0. Raise guidance_scale to 4–5.
• OOM at 1024 - drop to 768×768 or 832×832; the 1.3B runs on small GPUs but 1024² still adds up, especially in bf16.
• Weird anatomy - known weakness of this preview, per the author. It's trained to ~20% of budget; try different seeds, LoRA assistance, or lower steps before blaming your graph.
• Blank or stuck progress - check the Loader actually produced a pipe (the cache clears between some config changes) and that the diffusers version pin wasn't disturbed by another node pack.