bernini_r.md

Bernini-R — renderer only

← Back to main README

Bernini-R is the renderer-only Bernini model, fine-tuned from the Wan diffusion renderer. It skips the semantic-planning stage of the full Bernini pipeline, which makes it the recommended lightweight choice for simpler inference, renderer benchmarking, and scenarios where output consistency matters more than complex instruction following.

Checkpoint	Base	Notes
ByteDance/Bernini-R-Diffusers	Wan2.2-T2V-A14B	Recommended. Self-contained diffusers-format directory.
ByteDance/Bernini-R	Wan2.2-T2V-A14B	Separate high-/low-noise checkpoints; needs the Wan2.2 base download.
ByteDance/Bernini-R-1.3B-Diffusers	Wan2.1-1.3B	Lightweight variant; close to 14B on simple tasks (style transfer, subtitle/watermark removal, local editing), weaker on complex tasks such as human generation.

Benchmarks for the 14B and 1.3B variants are in the main README.

Download weights

Diffusers format (recommended)

The diffusers-format directory is self-contained: it includes the Wan base components plus the Bernini-R transformer / transformer_2 weights. Pass the directory directly as --config and do not pass --high_noise_ckpt / --low_noise_ckpt.

pip install -U "huggingface_hub"
hf download ByteDance/Bernini-R-Diffusers \
    --local-dir pretrained_models/Bernini-R-Diffusers

python infer_single_gpu.py --config pretrained_models/Bernini-R-Diffusers \
    --case assets/testcases/t2i/t2i.json --num_frames 1 --guidance_mode t2v_apg

The 1.3B release works the same way with ByteDance/Bernini-R-1.3B-Diffusers.

Separate checkpoints

This layout keeps the Wan2.2 base and the trained Bernini-R renderer weights separate. Use it only if you specifically need explicit high-noise / low-noise checkpoint paths.

pip install -U "huggingface_hub"
hf download Wan-AI/Wan2.2-T2V-A14B-Diffusers \
    --local-dir pretrained_models/Wan2.2-T2V-A14B-Diffusers
hf download ByteDance/Bernini-R \
    --local-dir pretrained_models/Bernini-R

Load the renderer config from configs/bernini_renderer_wan22 and pass both checkpoint paths (replace the file names with the actual safetensors in your download):

torchrun --nproc-per-node 8 infer_multi_gpu.py \
    --config configs/bernini_renderer_wan22 \
    --high_noise_ckpt pretrained_models/Bernini-R/<high-noise>.safetensors \
    --low_noise_ckpt pretrained_models/Bernini-R/<low-noise>.safetensors \
    --case assets/testcases/t2v/t2v.json

A matching config for the Wan2.1-1.3B base is provided at configs/bernini_renderer_wan21_1p3b.

Run

Make sure the environment is set up first — see Installation, which includes the required VeOmni install.

For single-GPU image tasks, use infer_single_gpu.py; for video tasks, use infer_multi_gpu.py with torchrun and --ulysses sequence parallelism:

# Single-GPU text-to-image
python infer_single_gpu.py --config pretrained_models/Bernini-R-Diffusers \
    --case assets/testcases/t2i/t2i.json --num_frames 1 --guidance_mode t2v_apg

# Multi-GPU video editing
torchrun --nproc-per-node 8 infer_multi_gpu.py \
    --config pretrained_models/Bernini-R-Diffusers --ulysses 8 \
    --case assets/testcases/v2v/v2v_case1.json --guidance_mode v2v_apg

Inputs are described by case files under assets/testcases/; see the case-file format.

Run scripts

scripts/bernini_r/ provides one script per task:

bash scripts/bernini_r/run_t2i.sh    # text-to-image
bash scripts/bernini_r/run_i2i.sh    # image editing
bash scripts/bernini_r/run_t2v.sh    # text-to-video
bash scripts/bernini_r/run_v2v.sh    # video editing
bash scripts/bernini_r/run_rv2v.sh   # reference + video editing
bash scripts/bernini_r/run_r2v.sh    # reference-to-video

The scripts use the diffusers layout and read these environment variables:

Variable	Default	Meaning
BERNINI_R_CONFIG	./pretrained_models/Bernini-R-Diffusers	model directory
CASE_PATH	a bundled example case	case JSON to run (single-case scripts)
NPROC_PER_NODE	8	number of processes (multi-GPU scripts)
ULYSSES	8	Ulysses sequence-parallel degree (multi-GPU scripts)

For the separate-checkpoint layout, replace --config with --config configs/bernini_renderer_wan22 --high_noise_ckpt <hi> --low_noise_ckpt <lo>.

See python infer_single_gpu.py --help for the full argument list.

Gradio demo

# Single GPU
python gradio_demo.py --config pretrained_models/Bernini-R-Diffusers --port 7860

# 8 GPUs, 8-way Ulysses sequence parallel
torchrun --nproc-per-node 8 gradio_demo.py --ulysses 8 \
    --config pretrained_models/Bernini-R-Diffusers --port 7860 --share

# Or the script launcher (diffusers layout, honors BERNINI_R_CONFIG)
bash scripts/bernini_r/run_gradio.sh

See the Gradio demo notes in the main README for the UI behavior and prompt-enhancer setup.