WorldFoundry
English中文

Bernini-R 14B

Method overview: an MLLM semantic planner plus a DiT renderer with segment-aware 3D RoPE on a unified multimodal sequence.

Bernini-R 14B from ByteDance: text-to-image and image-to-image; Runner parity pending; dedicated `worldfoundry-bernini-cu124`.

Edit this pageGitHubMarkdown

Bernini-R 14B is the renderer-only release of ByteDance's Bernini family: the 14B rendering model packaged without the 7B semantic planner. It keeps the same six-task surface (text-to-image, image editing, text/video/reference-to-video) but takes your inputs directly instead of planning first. WorldFoundry vendors the official renderer inference in-tree and loads ByteDance/Bernini-R-Diffusers.

Method overview: an MLLM semantic planner plus a DiT renderer with segment-aware 3D RoPE on a unified multimodal sequence.
Method overview: an MLLM semantic planner plus a DiT renderer with segment-aware 3D RoPE on a unified multimodal sequence.

The recorded paper is Bernini: Latent Semantic Planning for Video Diffusion (2026).

Project page: bernini-ai.github.io. Code: bytedance/Bernini.

This is not planner+renderer Bernini and not the single-GPU Bernini-R 1.3B.

Same paper: Bernini: Latent Semantic Planning for Video Diffusion. Weights: ByteDance/Bernini-R-Diffusers (apache-2.0). Image recipes are recorded as one-GPU; video recipes can turn on explicit Ulysses sequence parallelism. GPU artifact parity is pending. Status: Runner parity pending.

Build your run

Pipeline
BerniniWorldFoundryPipelineworldfoundry.pipelines.bernini.pipeline_bernini
Task
Environmentworldfoundry-bernini-cu124
DeviceCUDA 12.4
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  bernini-r-14b \
  --pipeline.task-profile t2v \
  --json

What it is for

Pick bernini-r-14b when you want this family's text-to-video line in WorldFoundry: a prompt in, a generated clip out. Use it as the catalog T2V start rather than an editor or a world-model stepper.

A single 14B rendering network from the Bernini two-stage design, run standalone. The manifest records that image recipes fit on one GPU while video recipes can use explicit Ulysses sequence parallelism for speed.

Catalog tasks include text-to-image, image-to-image, text-to-video, and video-to-video. Catalog aliases include bernini-r and bernini-r-diffusers. Documented modalities: text, image, and video in, image and video out.

Recorded evaluation entry points include VBench (Checks the renderer's raw T2V quality).

Runtime contract

Optional prompt, task_type, neg_prompt, num_frames, and height; artifacts generated_image, generated_video, and bernini-r-14b.mp4.

CLI task profiles: t2v, t2i, i2i, v2v, r2v, and rv2v. WorldFoundry binds BerniniWorldFoundryPipeline.

Family status: Runner parity pending.

Runner parity is pending_gpu_validation. Use task_type to select the task; stage the Bernini-R-Diffusers checkpoint (apache-2.0) locally first. For quick experiments prefer the 1.3B sibling; for the planner+renderer pipeline use the flagship bernini entry.

Official renderer inference is vendored in-tree; no external Bernini checkout is used.

Running in WorldFoundry

Launch worldfoundry-eval run bernini-r-14b --pipeline.task-profile t2v to sample the clip. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the dedicated environment worldfoundry-bernini-cu124 (Python 3.11, CUDA 12.4, torch==2.5.1+cu124).

Recorded weight repositories include ByteDance/Bernini-R-Diffusers and ByteDance/Bernini-R-Diffusers. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh bernini-r-14b. Add --download to fetch missing public checkpoints.

Recorded licenses include apache-2.0; status: Runner parity pending.