Krea Realtime 14B
What they did
Krea's realtime model, distilled from Wan 2.1 14B with Self-Forcing into an autoregressive generator. You can change prompts mid-generation, restyle on the fly and see first frames within about a second. It also streams video, webcam or canvas input for video-to-video.
Worth knowing
- Price: Open weights on Hugging Face (apache-2.0)
- License: apache-2.0
- Status: sourced. We have not run this ourselves yet (no "Tested" badge).
- Quality bar (computed by code): 5/5 (primary source ✓, reproducible ✓, new ✓, example output ✓, safe ✓)
The best new AI video workflows and prompts, weekly.
Related
▶ demo at sourceReplicate video collection
Replicate's curated collection of text-to-video models you can run by API, an alternative host to fal.
▶ demo at sourceComfyUI docs: video
Official ComfyUI tutorial for Wan 2.1: text-to-video with the 1.3B model and image-to-video with the 14B model, including where each model file goes.
▶ demo at sourceWan 2.5 Preview on fal
A preview of Wan 2.5 text-to-video on fal, with per-second prices that rise from 480p to 1080p. The page also accepts an audio file input.
▶ demo at sourceWan 2.2 on GitHub
Alibaba's official open-weights repo for Wan 2.2, the successor to 2.1, with code and model downloads for local runs.
▶ YouTube demoKeyframes to full character animation in ComfyUI (Wan 2.2 FLF)
The official ComfyUI tutorial for in-betweening: you draw key poses, each consecutive pair runs through Wan 2.2's first-last-frame model, and the clips are thinned, assembled in order and rendered as one animation.
▶ demo at sourceWan 2.1 on fal
Wan 2.1 image-to-video hosted on fal, with roughly 5–6 second clips at 480p or 720p and a flat price per video.