Still frames and live wallpapers in this album are generated locally with ComfyUI. Keyframes come from an Anima text-to-image graph; short motion loops are built with Wan-style first–last-frame video generation.

1. Text-to-image (Anima Aesthetic v1.1) — prompt in Danbooru / natural-language style with Qwen3 text encoder + Qwen-Image VAE. Base pass at 1344×768 (30 steps, CFG 4.5, er_sde).
2. Hi-res fix — latent upscale to 1920×1080, then a second sampler pass (≈20 steps, denoise ~0.45–0.50) to recover linework instead of naive resize.
3. Live wallpaper (Wan FLF2V / I2V) — use a selected still as start (and often end) frame for a seamless loop: subtle breathing, hair / fabric motion, light shifts; keep camera locked.
4. Curate & export — keep the best stills for the gallery grid; export short loops as live-wallpaper clips.

Stack: ComfyUI · Anima Aesthetic v1.1 · Qwen3 0.6B · Qwen-Image VAE · Wan 2.x video nodes.

Download ComfyUI workflow (JSON)