README.md

metadata

title: Alaya-EVOKE Text-to-Video
emoji: 🌍
colorFrom: yellow
colorTo: pink
sdk: gradio
sdk_version: 5.49.1
python_version: '3.10'
app_file: app.py
pinned: false
license: apache-2.0
short_description: Text-to-video world model (Evoke, 14B DiT, 3-step distilled)
startup_duration_timeout: 60m
models:
  - SII-YuanyangYin/Evoke
tags:
  - text-to-video
  - world-model
  - video-generation

Alaya-EVOKE · Text-to-Video

Prompt-only text-to-video demo of the released post-distillation checkpoint of SII-YuanyangYin/Evoke — Alaya-EVOKE: From Linear-Scaling Supervision to Endless World.

• 14B DiT world model + UMT5 text encoder + Wan VAE
• Sampled in 3 steps, no CFG (the distilled recipe) at 384×640 · 24 fps
• This is the model's t2v mode: no camera control (the pipeline runs warp-free), so the clip is generated from the prompt alone. One chunk ≈ 1.5 s of video.

The demo drives the authors' own inference engine (scripts/inference/infer_single.py) in-process, reproducing the infer_post_distill.sh MODE=t2v recipe exactly (STAGE2_STEPS=1 1 1, guidance_scale=1.0, vae_decode_type=persistent).

Note: the distilled checkpoints were trained on v2v conditioning, so t2v here is zero-shot extrapolation (as documented upstream) — expect it to work but not to be the model's trained mode.

Model & code © the Evoke authors, Apache-2.0.