diff --git a/README.md b/README.md
index 8cd8e9c..8cee834 100644
--- a/README.md
+++ b/README.md
@@ -23,12 +23,13 @@ WanGP supports the Wan (and derived models), Hunyuan Video and LTV Video models
 ### September 23 2025: WanGP v8.7 - Here Are Two New Contenders in the Vace Arena !
 
 So in today's release you will find two Wannabe Vace that covers each only a subset of Vace features but offers some interesting advantages:
-- **Wan 2.2 Animate**: this model is specialized in *Body Motion* and *Facial Motion tranfers*. It does that very well. You can use this model to either *Replace* a person in an in Video or *Animate* the person of your choice using an existing *Pose Video* (remember *Animate Anyone* ?). By default it will keep the original soundtrack. *Wan 2.2 Animate* seems to be under the hood a derived i2v model and should support the corresponding Loras Accelerators (for instance *FusioniX t2v*). Also as a WanGP exclusivity, you will find support for *Outpainting*.
+- **Wan 2.2 Animate**: this model is specialized in *Body Motion* and *Facial Motion transfers*. It does that very well. You can either *Replace* a person in a Video or *Animate* the person of your choice using an existing *Pose Video* (remember *Animate Anyone* ?). By default it will keep the original soundtrack. *Wan 2.2 Animate* seems to be under the hood a derived i2v model and should support the corresponding Loras Accelerators (for instance *FusioniX i2v*). Also as a WanGP exclusivity, you will find support for *Outpainting*.
 
-In order to use Wan 2.2 Animate you will need first to stop by the *Mat Anyone* embedded tool, to extract the Video Mask of the person from which you want to extract the motion.
+In order to use Wan 2.2 Animate you will need first to stop by the *Mat Anyone* embedded tool, to extract the *Video Mask* of the person from which you want to extract the motion.
 
-- **Lucy Edit**: this one claims to be a *Nano Banana* for Videos. Give it a video and asks it to change it (it is specialized in clothes changing) and voila ! The nice thing about it is that is it based on the *Wan 2.2 5B* model and therefore is very fast especially if you the *FastWan* finetune that is also part of the package. 
+- **Lucy Edit**: this one claims to be a *Nano Banana* for Videos. Give it a video and asks it to change it (it is specialized in clothes changing) and voila ! The nice thing about it is that is it based on the *Wan 2.2 5B* model and therefore is very fast especially if you the *FastWan* finetune that is also part of the package.
 
+*Update 8.71*: fixed Fast Lucy Edit that didnt contain the lora
 
 ### September 15 2025: WanGP v8.6 - Attack of the Clones
 
diff --git a/docs/INSTALLATION.md b/docs/INSTALLATION.md
index 9f66422..361f266 100644
--- a/docs/INSTALLATION.md
+++ b/docs/INSTALLATION.md
@@ -27,7 +27,7 @@ conda activate wan2gp
 ### Step 2: Install PyTorch
 
 ```shell
-# Install PyTorch 2.7.0 with CUDA 12.4
+# Install PyTorch 2.7.0 with CUDA 12.8
 pip install torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu128
 ```
 
diff --git a/requirements.txt b/requirements.txt
index e6c85fd..d6f75f7 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -52,7 +52,7 @@ matplotlib
 # Utilities
 ftfy
 piexif
-pynvml
+nvidia-ml-py 
 misaki
 
 # Optional / commented out
diff --git a/wgp.py b/wgp.py
index 43f5e80..d48c7fa 100644
--- a/wgp.py
+++ b/wgp.py
@@ -4859,7 +4859,7 @@ def generate_video(
         if repeat_no >= total_generation: break
         repeat_no +=1
         gen["repeat_no"] = repeat_no
-        src_video = src_video2 = src_mask = src_mask2 = src_faces = src_ref_images = src_ref_masks = None
+        src_video = src_video2 = src_mask = src_mask2 = src_faces = src_ref_images = src_ref_masks = sparse_video_image = None
         prefix_video = pre_video_frame = None
         source_video_overlap_frames_count = 0 # number of frames overalapped in source video for first window
         source_video_frames_count = 0  # number of frames to use in source video (processing starts source_video_overlap_frames_count frames before )