bernini-r-1.3b-diffusers-bf16

This repository contains a BF16 repack of ByteDance/Bernini-R-1.3B-Diffusers for local Apple Silicon inference with mlx-gen.

The source repository ships FP32 weights (~27 GiB) that inference runtimes cast to BF16 at load time. This repack stores the runtime dtypes directly, reducing the download to ~15.6 GiB (42% smaller) with no change in output:

• UMT5 text encoder: BF16, with the wo feed-forward projections stored in FP32 exactly as the runtime keeps them (_keep_in_fp32_modules).
• Renderer transformer: BF16, with the runtime FP32 keep-set (norms, scale-shift tables, and embedding layers) stored in FP32 exactly as the loader produces them.
• VAE: FP32, unchanged.
• Tokenizer, scheduler, and configs: unchanged.

The repository keeps the upstream Diffusers layout, so it also remains loadable by the official Bernini inference code.

Source Model

Original model: ByteDance/Bernini-R-1.3B-Diffusers.

This derivative follows the Apache 2.0 license of the source model.

Usage With mlx-gen

mlxgen download --model bernini-r-1.3b-bf16

mlxgen generate \
  --model bernini-r-1.3b-bf16 \
  --reference-image subject.png \
  --prompt "Bring the subject from image0 to life in a fixed medium shot" \
  --width 848 --height 480 --frames 81 --fps 16 --steps 40 \
  --seed 42 --output referenced.mp4

Bernini-R supports reference-to-video (1-8 reference images), reference-guided video editing (--video plus references), and prompt-guided video editing (--video only). See the mlx-gen Bernini documentation for workflows, guidance defaults, and task-specific recipes.

Fidelity

Tensors are bit-exact casts of the pinned source revision (ff4c5d4d2d31365c2ffeb30e9753065ee18f58ce): every tensor equals the value the mlx-gen loader produces from the FP32 original, and generation output was verified bit-identical to the source repository on image and video use cases at the same settings and seed (max pixel diff 0).