rzgar / Bernini-R-n4w

Bernini-R-n4w

This merge handles male and female anatomy well. Results become significantly stronger and more consistent when guided by reference images or reference videos.

Success Rates (based on internal testing):

• I2V / Reference Video-to-Video : Over 75% good results in a single generation
• Text-to-Video (T2V) : Above 65% when using detailed prompts (quality varies with prompt strength)

Prompting Advice

Bernini-R is noticeably more complex than standard Wan2.2 models, so the short/simple prompts many people use with Wan2.2 often don’t perform well here.

• Text-to-Video (T2V) : Prompts need to be detailed and descriptive . In some cases, running a second generation was required to get a strong result.
• Image-to-Video with reference video : One generation was enough in roughly 7 out of 10 tests.

Tip – Refiner for Difficult Scenes

For tricky or complex scenes, using the low-noise expert from another Wan2.2 merge (I used SmoothMix I2V V1 low-noise ) as a refiner often further improves anatomy and fine details if the initial video isn’t too messy.

LightX2V LoRA

• wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors
• wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors

Text Encoders

• umt5_xxl_fp8_e4m3fn_scaled.safetensors

VAE

• wan_2.1_vae.safetensors

Clip Vision

• clip_vision_h.safetensors