Krea realtime video (#1011)

* krea-realtime-video

* Add Krea real-time video inference and training support

* Delete .gitignore

* update README

* update README

---------

Co-authored-by: Artiprocher <wangye87v5@hotmail.com>
Co-authored-by: Jintao Huang <huangjintao.hjt@alibaba-inc.com>
Co-authored-by: Zhongjie Duan <35051019+Artiprocher@users.noreply.github.com>

examples/wanvideo/model_inference/krea-realtime-video.py

@@ -0,0 +1,25 @@
+import torch
+from diffsynth import save_video
+from diffsynth.pipelines.wan_video_new import WanVideoPipeline, ModelConfig
+
+
+pipe = WanVideoPipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="krea/krea-realtime-video", origin_file_pattern="krea-realtime-video-14b.safetensors", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="Wan2.1_VAE.pth", offload_device="cpu"),
+    ],
+)
+pipe.enable_vram_management()
+
+# Text-to-video
+video = pipe(
+    prompt="a cat sitting on a boat",
+    num_inference_steps=6, num_frames=81,
+    seed=0, tiled=True,
+    cfg_scale=1,
+    sigma_shift=20,
+)
+save_video(video, "video1.mp4", fps=15, quality=5)