Image-to-Image
Diffusers
lora
template:diffusion-lora
text-to-video
image-to-video
video-to-video
lightx2v
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("ByteDance/Bernini-R", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("rzgar/Bernini-R-LightX2V-4step-loras")
prompt = "Turn this cat into a dog"
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
image = pipe(image=input_image, prompt=prompt).images[0]Bernini-R LightX2V LoRAs (4-step)
## Model descriptionWhy not lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16?
The ComfyUI Bernini-R workflow uses a single Wan 2.1 T2V lightx2v LoRA, reused at 3.0 (high) and 1.5 (low).
In practice that setup often shows:
- Blurry fingers on movement
- Laggy or glitchy motion when things move quickly
These LightX2V LoRAs behave like the Wan 2.2 4-step stack , clean motion, sharp hands , at 1.0 strength on both stages.
Files
| File | Use on |
|---|---|
| `Bernini-R_LightX2V_high_noise.safetensors` | Bernini-R high noise |
| `Bernini-R_LightX2V_low_noise.safetensors` | Bernini-R low noise |
Strength: 1.0 on both (same as the official Wan 2.2 4-step workflow).
Steps: 4
Sampler: dpmpp_2m_sde
Scheduler: sgm_uniform
Credits
Download model
Download them in the Files & versions tab.
- Downloads last month
- -
Model tree for rzgar/Bernini-R-LightX2V-4step-loras
Base model
ByteDance/Bernini-R