mlx-community/Bernini-v2-bf16

Bernini-v2 (bf16, MLX)

Apple-MLX conversion of ByteDance/Bernini-Diffusers-v2 (revision 399cf6a) — the full unified Bernini: MLLM semantic planner + dual-expert Wan2.2-A14B DiT renderer. Converted 2026-08-18. Apache-2.0, same as upstream.

Contents

File	Component	Notes
high_noise_model.safetensors	Wan2.2-A14B high-noise expert (bf16)	retrained vs Bernini-R (co-trained with the planner) — not interchangeable with mlx-community/Bernini-R-bf16
low_noise_model.safetensors	Wan2.2-A14B low-noise expert (bf16)	ditto
mllm/	Qwen2.5-VL-7B semantic planner (bf16, HF layout)	Bernini-trained weights (scratch_mllm), not stock Qwen; configs + tokenizer from upstream
vit_decoder.safetensors	DiffLoss_FM flow-match head (bf16)	SimpleMLPAdaLN, width 4096, depth 16
planner_glue.safetensors	MLPConnector + mask_tokens	keys verbatim upstream
t5_encoder.safetensors	umT5-XXL (bf16)	bit-identical to the stock Wan2.2 encoder (verified vs upstream)
vae.safetensors	16-ch WanVAE	bit-identical to stock Wan2.2 (verified vs upstream)
config.json	wan-core runtime config (dual-expert A14B)
conversion.json	conversion provenance

Conversion notes

• Experts: upstream fp32 masters → diffusers→original-Wan key premap → mlx-video sanitize → bf16. Key set verified bijective against the established Bernini-R MLX layout; value probes bit-exact (RNE) against the fp32 masters.
• The planner (mllm.*) is saved in standard HF Qwen2.5-VL layout for direct consumption by MLX Qwen2.5-VL loaders.
• The upstream in-checkpoint fp32 T5 copy was skipped; the standalone bf16 encoder (verified bit-identical) is shipped instead.

Usage

The renderer is drop-in for the Bernini-R MLX stack (same layout as mlx-community/Bernini-R-bf16) — e.g. bernini-r-mlx pipeline_mlx.t2v/t2i, or the Swift bernini-r-mlx-swift/wan-core stack.

The planner plane (mllm / vit_decoder / connector / mask_tokens) implements the MaskGIT-style semantic planning of the Bernini paper (arXiv 2605.22344); a Swift-MLX planner integration is in progress in bernini-r-mlx-swift. Until then these files carry the released weights for downstream use.

Provenance & license

Upstream: ByteDance/Bernini-Diffusers-v2 (Apache-2.0). All credit for the model to the Bernini authors — see the Bernini repository and paper. This conversion changes dtype/layout only (plus the key renames described above); no weights were fine-tuned.