[Feat] Add load lora for animate (#433)

configs/wan22/wan_animate_lora.json

+{
+    "infer_steps": 4,
+    "target_video_length": 77,
+    "text_len": 512,
+    "target_height": 720,
+    "target_width": 1280,
+    "self_attn_1_type": "flash_attn3",
+    "cross_attn_1_type": "flash_attn3",
+    "cross_attn_2_type": "flash_attn3",
+    "adapter_attn_type": "flash_attn3",
+    "sample_shift": 5.0,
+    "sample_guide_scale": 1.0,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "src_pose_path": "../save_results/animate/process_results/src_pose.mp4",
+    "src_face_path": "../save_results/animate/process_results/src_face.mp4",
+    "src_ref_images": "../save_results/animate/process_results/src_ref.png",
+    "refert_num": 1,
+    "replace_flag": false,
+    "fps": 30,
+    "lora_configs": [
+        {
+        "path": "lightx2v_I2V_14B_480p_cfg_step_distill_rank32_bf16.safetensors",
+        "strength": 1.0
+        }
+    ]
+}

lightx2v/models/runners/wan/wan_animate_runner.py

@@ -17,6 +17,7 @@ except ImportError:
 from lightx2v.models.input_encoders.hf.animate.face_encoder import FaceEncoder
 from lightx2v.models.input_encoders.hf.animate.motion_encoder import Generator
 from lightx2v.models.networks.wan.animate_model import WanAnimateModel
+from lightx2v.models.networks.wan.lora_adapter import WanLoraWrapper
 from lightx2v.models.runners.wan.wan_runner import WanRunner
 from lightx2v.server.metrics import monitor_cli
 from lightx2v.utils.envs import *
@@ -391,6 +392,17 @@ class WanAnimateRunner(WanRunner):
             self.config,
             self.init_device,
         )
+
+        if self.config.get("lora_configs") and self.config.lora_configs:
+            assert not self.config.get("dit_quantized", False)
+            lora_wrapper = WanLoraWrapper(model)
+            for lora_config in self.config.lora_configs:
+                lora_path = lora_config["path"]
+                strength = lora_config.get("strength", 1.0)
+                lora_name = lora_wrapper.load_lora(lora_path)
+                lora_wrapper.apply_lora(lora_name, strength)
+                logger.info(f"Loaded LoRA: {lora_name} with strength: {strength}")
+
         motion_encoder, face_encoder = self.load_encoders()
         model.set_animate_encoders(motion_encoder, face_encoder)
         return model

scripts/wan22/run_wan22_animate_lora.sh

+#!/bin/bash
+
+# set path and first
+lightx2v_path=
+model_path=
+video_path=
+refer_path=
+
+export CUDA_VISIBLE_DEVICES=0
+
+# set environment variables
+source ${lightx2v_path}/scripts/base/base.sh
+
+# process
+python ${lightx2v_path}/tools/preprocess/preprocess_data.py \
+    --ckpt_path ${model_path}/process_checkpoint \
+    --video_path $video_path  \
+    --refer_path $refer_path \
+    --save_path ${lightx2v_path}/save_results/animate/process_results \
+    --resolution_area 1280 720 \
+    --retarget_flag \
+
+python -m lightx2v.infer \
+--model_cls wan2.2_animate \
+--task animate \
+--model_path $model_path \
+--config_json ${lightx2v_path}/configs/wan22/wan_animate_lora.json \
+--prompt "视频中的人在做动作" \
+--negative_prompt "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" \
+--save_result_path ${lightx2v_path}/save_results/output_lightx2v_wan22_animate_lora.mp4