rzgar/Bernini-v2-ComfyUI

Quantized and full-precision (fp32) extraction of every component from the sharded bernini/ bundle. Tensor keys are preserved verbatim (including their original prefixes).

Quantized

• NVFP4
• FP8_scaled

Components

• mllm fp16/fp8
• vit)decoder_bf16
• connector
• mask_tokens
• umt5_xxl_fp16/fp8
• vae_fp32

LightX2V 4Step

• Bernini-R_LightX2V_high_noise
• Bernini-R_LightX2V_low_noise

---

Original bundel

File	Component	Tensors	Size	Original prefix
connector.safetensors	MLP connector (planner→renderer)	12	0.28 GB	connector.
mask_tokens.safetensors	mask-token embedding (4096 × 3584)	1	0.06 GB	mask_tokens
vit_decoder.safetensors	vision decoder	140	5.75 GB	vit_decoder.
mllm.safetensors	Qwen2.5-VL planner (fine-tuned)	729	33.17 GB	mllm.
t5_text_encoder_fp32.safetensors	UMT5-XXL text encoder	243	26.92 GB	t5_text_encoder.
high_noise_fp32.safetensors	Wan2.2 high-noise DiT	1095	57.15 GB	diff_dec.
low_noise_fp32.safetensors	Wan2.2 low-noise DiT	1095	57.15 GB	diff_dec_low.