README.md

metadata

title: Bernini Diffusers v2
emoji: 🗿
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 6.15.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Reference-to-video with ByteDance Bernini-Diffusers-v2
python_version: '3.12'
startup_duration_timeout: 2h
models:
  - ByteDance/Bernini-Diffusers-v2

Bernini-Diffusers-v2 — reference-to-video

Give it a handful of reference images (a subject, an outfit, a prop, a scene…) and a prompt that points at them as image0, image1, … Bernini's Qwen2.5-VL planner reads the references together with the instruction and plans a target visual embedding with a flow-matching head; the Wan2.2-A14B MoE renderer (two 14 B DiTs, high-noise + low-noise) turns that plan into a video.

• Model: ByteDance/Bernini-Diffusers-v2
• Code: bytedance/Bernini

What this Space runs

The r2v (reference-to-video) task, matching the authors' scripts/bernini_v2/run_r2v.sh one-for-one: guidance_mode=vae_txt_vit_wapg, omega_txt=4.5, omega_tgt=1.5, omega_img=3.0, omega_vid=1.0, omega_scale=0.75, planning_step=50, vit_denoising_step=1, vit_txt_cfg=1.2, vit_img_cfg=1.0, flow_shift=5.0, max_image_size=842, 16 fps, and the same system / negative prompt.

The only deviation is the default clip length and step count (33 frames / 16 steps instead of 81 / 40), so a generation fits inside a single ZeroGPU slot — both are sliders under Advanced settings. At the defaults a video takes about 4 minutes.

The released checkpoint is fp32 (~180 GB); it is loaded in bf16, which is the dtype the reference pipeline computes in anyway.

Credits

The bernini/ package and the veomni/ subset shipped alongside app.py are vendored from bytedance/Bernini and ByteDance-Seed/VeOmni (v0.1.11), both Apache-2.0, because both declare requires-python ranges that exclude this runtime.

The example reference images in examples/ are the authors' own r2v test case assets from bytedance/Bernini (Apache-2.0).