movie import fix

Package

README.md

@@ -2,32 +2,48 @@
 
 Latent blending enables video transitions with incredible smoothness between prompts, computed within seconds. Powered by [stable diffusion XL](https://stability.ai/stable-diffusion), this method involves specific mixing of intermediate latent representations to create a seamless transition – with users having the option to fully customize the transition directly in high-resolution. The new version also supports SDXL Turbo, allowing to generate transitions faster than they are typically played back!
 
-
 ```python
+import torch
+from diffusers import AutoPipelineForText2Image
+from latentblending.blending_engine import BlendingEngine
+from latentblending.diffusers_holder import DiffusersHolder
+
 pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16").to("cuda")
-dh = DiffusersHolder(pipe)
-lb = LatentBlending(dh)
-lb.set_prompt1("photo of underwater landscape, fish, und the sea, incredible detail, high resolution")
-lb.set_prompt2("rendering of an alien planet, strange plants, strange creatures, surreal")
-lb.set_negative_prompt("blurry, ugly, pale")
+be = BlendingEngine(pipe)
+be.set_prompt1("photo of underwater landscape, fish, und the sea, incredible detail, high resolution")
+be.set_prompt2("rendering of an alien planet, strange plants, strange creatures, surreal")
+be.set_negative_prompt("blurry, ugly, pale")
 
 # Run latent blending
-lb.run_transition()
+be.run_transition()
 
 # Save movie
-lb.write_movie_transition('movie_example1.mp4', duration_transition=12)
-
+be.write_movie_transition('movie_example1.mp4', duration_transition=12)
 ```
+
+# Installation
+```commandline
+pip install git+https://github.com/lunarring/latentblending
+```
+
+# Extra speedup with stable_fast compile
+Install https://github.com/chengzeyi/stable-fast
+
+Then enable pipe compilation by setting *do_compile=True*
+```python
+be = BlendingEngine(pipe, do_compile=True)
+```
+
 ## Gradio UI
 Coming soon again :)
 
 ## Example 1: Simple transition
 ![](example1.jpg)
-To run a simple transition between two prompts, run `example1_standard.py`
+To run a simple transition between two prompts, see `examples/single_trans.py`
 
 ## Example 2: Multi transition
-To run multiple transition between K prompts, resulting in a stitched video, run `example2_multitrans.py`.
-[View a longer example video here.](https://vimeo.com/789052336/80dcb545b2)
+To run multiple transition between K prompts, resulting in a stitched video, see `examples/multi_trans.py`.
+[View a longer example video here.](https://youtu.be/RLF-yW5dR_Q) 
 
 
 # Customization
@@ -35,19 +51,19 @@ To run multiple transition between K prompts, resulting in a stitched video, run
 ### Change the height/width
 ```python 
 size_output = (1024, 768)
-lb.set_dimensions(size_output)
+be.set_dimensions(size_output)
 ```
 
 ### Change the number of diffusion steps (set_num_inference_steps)
 ```python
-lb.set_num_inference_steps(50)
+be.set_num_inference_steps(50)
 ```
 For SDXL this is set as default=30, for SDXL Turbo a value of 4 is taken.
 
 
 ### Change the guidance scale
 ```python
-lb.set_guidance_scale(3.0)
+be.set_guidance_scale(3.0)
 ```
 For SDXL this is set as default=4.0, for SDXL Turbo a value of 0 is taken.
 
@@ -55,7 +71,7 @@ For SDXL this is set as default=4.0, for SDXL Turbo a value of 0 is taken.
 ```python
 depth_strength = 0.5
 nmb_max_branches = 15
-lb.set_branching(depth_strength=depth_strength, t_compute_max_allowed=None, nmb_max_branches=None)
+be.set_branching(depth_strength=depth_strength, t_compute_max_allowed=None, nmb_max_branches=None)
 ```
 * depth_strength: The strength of the diffusion iterations determines when the blending process will begin. A value close to zero results in more creative and intricate outcomes, while a value closer to one indicates a simpler alpha blending. However, low values may also bring about the introduction of additional objects and motion.
 * t_compute_max_allowed:  maximum time allowed for computation. Higher values give better results but take longer. Either provide t_compute_max_allowed or nmb_max_branches. Does not work for SDXL Turbo.
@@ -66,7 +82,7 @@ You can find the [most relevant parameters here.](parameters.md)
 
 ### Change guidance scale
 ```python 
-lb.set_guidance_scale(5.0)
+be.set_guidance_scale(5.0)
 ```
 
 ### Crossfeeding to the last image.
@@ -76,7 +92,7 @@ Cross-feeding latents is a key feature of latent blending. Here, you can set how
 crossfeed_power = 0.5 # 50% of the latents in the last branch are copied from branch1
 crossfeed_range = 0.7 # The crossfeed is active until 70% of num_iteration, then switched off
 crossfeed_decay = 0.2 # The power of the crossfeed decreases over diffusion iterations, here it would be 0.5*0.2=0.1 in the end of the range.
-lb.set_branch1_crossfeed(crossfeed_power, crossfeed_range, crossfeed_decay)
+be.set_branch1_crossfeed(crossfeed_power, crossfeed_range, crossfeed_decay)
 ```
 
 ### Crossfeeding to all transition images
@@ -86,16 +102,10 @@ Here, you can set how much the parent branches influence the mixed one. In the a
 crossfeed_power = 0.5 # 50% of the latents in the last branch are copied from the parents
 crossfeed_range = 0.7 # The crossfeed is active until 70% of num_iteration, then switched off
 crossfeed_decay = 0.2 # The power of the crossfeed decreases over diffusion iterations, here it would be 0.5*0.2=0.1 in the end of the range.
-lb.set_parental_crossfeed(crossfeed_power, crossfeed_range, crossfeed_decay)
+be.set_parental_crossfeed(crossfeed_power, crossfeed_range, crossfeed_decay)
 ```
 
 
-# Installation
-#### Packages
-```commandline
-pip install -r requirements.txt
-```
-
 # How does latent blending work?
 ## Method
 ![](animation.gif)
@@ -104,9 +114,9 @@ In the figure above, a diffusion tree is illustrated. The diffusion steps are re
 
 The concrete parameters for the transition above would be:
 ```
-lb.set_branch1_crossfeed(crossfeed_power=0.8, crossfeed_range=0.6, crossfeed_decay=0.4)
-lb.set_parental_crossfeed(crossfeed_power=0.8, crossfeed_range=0.8, crossfeed_decay=0.2)
-imgs_transition = lb.run_transition(num_inference_steps=10, depth_strength=0.2, nmb_max_branches=7)
+be.set_branch1_crossfeed(crossfeed_power=0.8, crossfeed_range=0.6, crossfeed_decay=0.4)
+be.set_parental_crossfeed(crossfeed_power=0.8, crossfeed_range=0.8, crossfeed_decay=0.2)
+imgs_transition = be.run_transition(num_inference_steps=10, depth_strength=0.2, nmb_max_branches=7)
 ```
 
 ## Perceptual aspects
@@ -124,6 +134,7 @@ With latent blending, we can create transitions that appear to defy the laws of
 * Inpaint support dropped (as it only makes sense for a single transition)
 
 # Coming soon...
+- [ ] MacOS support
 - [ ] Gradio interface
 - [ ] Huggingface Space
 - [ ] Controlnet

examples/multi_trans.py

@@ -1,33 +1,42 @@
 import torch
 import warnings
-from blending_engine import BlendingEngine
-from diffusers_holder import DiffusersHolder
 from diffusers import AutoPipelineForText2Image
-from movie_util import concatenate_movies
+from latentblending.movie_util import concatenate_movies
+from latentblending.blending_engine import BlendingEngine
+import numpy as np
 torch.set_grad_enabled(False)
 torch.backends.cudnn.benchmark = False
 warnings.filterwarnings('ignore')
 
 # %% First let us spawn a stable diffusion holder. Uncomment your version of choice.
-pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
-pipe.to('cuda')
-dh = DiffusersHolder(pipe)
+pretrained_model_name_or_path = "stabilityai/stable-diffusion-xl-base-1.0"
+# pretrained_model_name_or_path = "stabilityai/sdxl-turbo"
 
+pipe = AutoPipelineForText2Image.from_pretrained(pretrained_model_name_or_path, torch_dtype=torch.float16, variant="fp16")
+pipe.to('cuda')
+be = BlendingEngine(pipe, do_compile=True)
+be.set_negative_prompt("blurry, pale, low-res, lofi")
 # %% Let's setup the multi transition
 fps = 30
 duration_single_trans = 10
+be.set_dimensions((1024, 1024))
+nmb_prompts = 20
+
+
+
+
 
 # Specify a list of prompts below
+#%%
+
 list_prompts = []
-list_prompts.append("Photo of a house, high detail")
-list_prompts.append("Photo of an elephant in african savannah")
-list_prompts.append("photo of a house, high detail")
-
-
-# You can optionally specify the seeds
-list_seeds = [95437579, 33259350, 956051013]
-fp_movie = 'movie_example2.mp4'
-be = BlendingEngine(dh)
+list_prompts.append("high resolution ultra 8K image with lake and forest")
+list_prompts.append("strange and alien desolate lanscapes 8K")
+list_prompts.append("ultra high res psychedelic skyscraper city landscape 8K unreal engine")
+#%%
+fp_movie = f'surreal_nmb{len(list_prompts)}.mp4'
+# Specify the seeds
+list_seeds = np.random.randint(0, np.iinfo(np.int32).max, len(list_prompts))
 
 list_movie_parts = []
 for i in range(len(list_prompts) - 1):

examples/single_trans.py

@@ -1,8 +1,7 @@
 import torch
 import warnings
-from blending_engine import BlendingEngine
-from diffusers_holder import DiffusersHolder
 from diffusers import AutoPipelineForText2Image
+from latentblending.blending_engine import BlendingEngine
 
 warnings.filterwarnings('ignore')
 torch.set_grad_enabled(False)
@@ -12,9 +11,7 @@ torch.backends.cudnn.benchmark = False
 pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
 pipe.to("cuda")
 
-dh = DiffusersHolder(pipe)
-
-be = BlendingEngine(dh)
+be = BlendingEngine(pipe)
 be.set_prompt1("photo of underwater landscape, fish, und the sea, incredible detail, high resolution")
 be.set_prompt2("rendering of an alien planet, strange plants, strange creatures, surreal")
 be.set_negative_prompt("blurry, ugly, pale")

latentblending/blending_engine.py

@@ -5,10 +5,12 @@ import warnings
 import time
 from tqdm.auto import tqdm
 from PIL import Image
-from latentblending.movie_util import MovieSaver
 from typing import List, Optional
 import lpips
-from latentblending.utils import interpolate_spherical, interpolate_linear, add_frames_linear_interp, yml_load, yml_save
+import platform
+from latentblending.diffusers_holder import DiffusersHolder
+from latentblending.utils import interpolate_spherical, interpolate_linear, add_frames_linear_interp
+from lunar_tools import MovieSaver, fill_up_frames_linear_interpolation
 warnings.filterwarnings('ignore')
 torch.backends.cudnn.benchmark = False
 torch.set_grad_enabled(False)
@@ -17,12 +19,15 @@ torch.set_grad_enabled(False)
 class BlendingEngine():
     def __init__(
             self,
-            dh: None,
+            pipe: None,
+            do_compile: bool = False,
             guidance_scale_mid_damper: float = 0.5,
             mid_compression_scaler: float = 1.2):
         r"""
         Initializes the latent blending class.
         Args:
+            pipe: diffusers pipeline (SDXL)
+            do_compile: compile pipeline for faster inference using stable fast
             guidance_scale_mid_damper: float = 0.5
                 Reduces the guidance scale towards the middle of the transition.
                 A value of 0.5 would decrease the guidance_scale towards the middle linearly by 0.5.
@@ -35,7 +40,8 @@ class BlendingEngine():
             and guidance_scale_mid_damper <= 1.0, \
             f"guidance_scale_mid_damper neees to be in interval (0,1], you provided {guidance_scale_mid_damper}"
 
-        self.dh = dh
+    
+        self.dh = DiffusersHolder(pipe)
         self.device = self.dh.device
         self.set_dimensions()
 
@@ -64,6 +70,9 @@ class BlendingEngine():
         self.multi_transition_img_first = None
         self.multi_transition_img_last = None
         self.dt_unet_step = 0
+        if platform.system() == "Darwin":
+            self.lpips = lpips.LPIPS(net='alex')
+        else:
             self.lpips = lpips.LPIPS(net='alex').cuda(self.device)
 
         self.set_prompt1("")
@@ -76,13 +85,23 @@ class BlendingEngine():
         self.benchmark_speed()
         self.set_branching()
         
+        if do_compile:
+            print("starting compilation")
+            from sfast.compilers.diffusion_pipeline_compiler import (compile, CompilationConfig)
+            self.dh.pipe.enable_xformers_memory_efficient_attention()
+            config = CompilationConfig.Default()
+            config.enable_xformers = True
+            config.enable_triton = True
+            config.enable_cuda_graph = True
+            self.dh.pipe = compile(self.dh.pipe, config)
+        
         
         
     def benchmark_speed(self):
         """
         Measures the time per diffusion step and for the vae decoding
         """
-        
+        print("starting speed benchmark...")
         text_embeddings = self.dh.get_text_embedding("test")
         latents_start = self.dh.get_noise(np.random.randint(111111))
         # warmup
@@ -96,6 +115,7 @@ class BlendingEngine():
         t0 = time.time()
         img = self.dh.latent2image(list_latents[-1])
         self.dt_vae = time.time() - t0
+        print(f"time per unet iteration: {self.dt_unet_step} time for vae: {self.dt_vae}")
 
     def set_dimensions(self, size_output=None):
         r"""
@@ -268,7 +288,7 @@ class BlendingEngine():
             if t_compute_max_allowed is None and nmb_max_branches is None:
                 t_compute_max_allowed = 20
             elif t_compute_max_allowed is not None and nmb_max_branches is not None:
-                raise ValueErorr("Either specify t_compute_max_allowed or nmb_max_branches")
+                raise ValueError("Either specify t_compute_max_allowed or nmb_max_branches")
             
             self.list_idx_injection, self.list_nmb_stems = self.get_time_based_branching(depth_strength, t_compute_max_allowed, nmb_max_branches)    
 
@@ -676,7 +696,7 @@ class BlendingEngine():
         """
 
         # Let's get more cheap frames via linear interpolation (duration_transition*fps frames)
-        imgs_transition_ext = add_frames_linear_interp(self.tree_final_imgs, duration_transition, fps)
+        imgs_transition_ext = fill_up_frames_linear_interpolation(self.tree_final_imgs, duration_transition, fps)
 
         # Save as MP4
         if os.path.isfile(fp_movie):
@@ -686,12 +706,6 @@ class BlendingEngine():
             ms.write_frame(img)
         ms.finalize()
 
-    def save_statedict(self, fp_yml):
-        # Dump everything relevant into yaml
-        imgs_transition = self.tree_final_imgs
-        state_dict = self.get_state_dict()
-        state_dict['nmb_images'] = len(imgs_transition)
-        yml_save(fp_yml, state_dict)
 
     def get_state_dict(self):
         state_dict = {}
@@ -813,14 +827,18 @@ if __name__ == "__main__":
     from diffusers import AutoencoderTiny
     # pretrained_model_name_or_path = "stabilityai/stable-diffusion-xl-base-1.0"
     pretrained_model_name_or_path = "stabilityai/sdxl-turbo"
+    pipe = DiffusionPipeline.from_pretrained(pretrained_model_name_or_path)
     
     
-    pipe = DiffusionPipeline.from_pretrained(pretrained_model_name_or_path, torch_dtype=torch.float16, variant="fp16")
+    # pipe.to("mps")
     pipe.to("cuda")
-    pipe.vae = AutoencoderTiny.from_pretrained('madebyollin/taesdxl', torch_device='cuda', torch_dtype=torch.float16)
-    pipe.vae = pipe.vae.cuda()
+    
+    # pipe.vae = AutoencoderTiny.from_pretrained('madebyollin/taesdxl', torch_device='cuda', torch_dtype=torch.float16)
+    # pipe.vae = pipe.vae.cuda()
 
     dh = DiffusersHolder(pipe)
+    
+    xxx
     # %% Next let's set up all parameters
     prompt1 = "photo of underwater landscape, fish, und the sea, incredible detail, high resolution"
     prompt2 = "rendering of an alien planet, strange plants, strange creatures, surreal"
@@ -829,19 +847,20 @@ if __name__ == "__main__":
     duration_transition = 12  # In seconds
 
     # Spawn latent blending
-    lb = LatentBlending(dh)
-    lb.set_prompt1(prompt1)
-    lb.set_prompt2(prompt2)
-    lb.set_negative_prompt(negative_prompt)
+    be = BlendingEngine(dh)
+    be.set_prompt1(prompt1)
+    be.set_prompt2(prompt2)
+    be.set_negative_prompt(negative_prompt)
 
     # Run latent blending
     t0 = time.time()
-    lb.run_transition(fixed_seeds=[420, 421])
+    be.run_transition(fixed_seeds=[420, 421])
     dt = time.time() - t0
+    print(f"dt = {dt}")
 
     # Save movie
     fp_movie = f'test.mp4'
-    lb.write_movie_transition(fp_movie, duration_transition)
+    be.write_movie_transition(fp_movie, duration_transition)
     
 
 

latentblending/movie_util.py

@@ -1,301 +0,0 @@
-# Copyright 2022 Lunar Ring. All rights reserved.
-# Written by Johannes Stelzer, email stelzer@lunar-ring.ai twitter @j_stelzer
-
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import subprocess
-import os
-import numpy as np
-from tqdm import tqdm
-import cv2
-from typing import List
-import ffmpeg  # pip install ffmpeg-python. if error with broken pipe: conda update ffmpeg
-
-
-class MovieSaver():
-    def __init__(
-            self,
-            fp_out: str,
-            fps: int = 24,
-            shape_hw: List[int] = None,
-            crf: int = 21,
-            codec: str = 'libx264',
-            preset: str = 'fast',
-            pix_fmt: str = 'yuv420p',
-            silent_ffmpeg: bool = True):
-        r"""
-        Initializes movie saver class - a human friendly ffmpeg wrapper.
-        After you init the class, you can dump numpy arrays x into moviesaver.write_frame(x).
-        Don't forget toi finalize movie file with moviesaver.finalize().
-        Args:
-            fp_out: str
-                Output file name. If it already exists, it will be deleted.
-            fps: int
-                Frames per second.
-            shape_hw: List[int, int]
-                Output shape, optional argument. Can be initialized automatically when first frame is written.
-            crf: int
-                ffmpeg doc: the range of the CRF scale is 0–51, where 0 is lossless
-                (for 8 bit only, for 10 bit use -qp 0), 23 is the default, and 51 is worst quality possible.
-                A lower value generally leads to higher quality, and a subjectively sane range is 17–28.
-                Consider 17 or 18 to be visually lossless or nearly so;
-                it should look the same or nearly the same as the input but it isn't technically lossless.
-                The range is exponential, so increasing the CRF value +6 results in
-                roughly half the bitrate / file size, while -6 leads to roughly twice the bitrate.
-            codec: int
-                Number of diffusion steps. Larger values will take more compute time.
-            preset: str
-                Choose between ultrafast, superfast, veryfast, faster, fast, medium, slow, slower, veryslow.
-                ffmpeg doc: A preset is a collection of options that will provide a certain encoding speed
-                to compression ratio. A slower preset will provide better compression
-                (compression is quality per filesize).
-                This means that, for example, if you target a certain file size or constant bit rate,
-                you will achieve better quality with a slower preset. Similarly, for constant quality encoding,
-                you will simply save bitrate by choosing a slower preset.
-            pix_fmt: str
-                Pixel format. Run 'ffmpeg -pix_fmts' in your shell to see all options.
-            silent_ffmpeg: bool
-                Surpress the output from ffmpeg.
-        """
-        if len(os.path.split(fp_out)[0]) > 0:
-            assert os.path.isdir(os.path.split(fp_out)[0]), "Directory does not exist!"
-
-        self.fp_out = fp_out
-        self.fps = fps
-        self.crf = crf
-        self.pix_fmt = pix_fmt
-        self.codec = codec
-        self.preset = preset
-        self.silent_ffmpeg = silent_ffmpeg
-
-        if os.path.isfile(fp_out):
-            os.remove(fp_out)
-
-        self.init_done = False
-        self.nmb_frames = 0
-        if shape_hw is None:
-            self.shape_hw = [-1, 1]
-        else:
-            if len(shape_hw) == 2:
-                shape_hw.append(3)
-            self.shape_hw = shape_hw
-            self.initialize()
-
-        print(f"MovieSaver initialized. fps={fps} crf={crf} pix_fmt={pix_fmt} codec={codec} preset={preset}")
-
-    def initialize(self):
-        args = (
-            ffmpeg
-            .input('pipe:', format='rawvideo', pix_fmt='rgb24', s='{}x{}'.format(self.shape_hw[1], self.shape_hw[0]), framerate=self.fps)
-            .output(self.fp_out, crf=self.crf, pix_fmt=self.pix_fmt, c=self.codec, preset=self.preset)
-            .overwrite_output()
-            .compile()
-        )
-        if self.silent_ffmpeg:
-            self.ffmpg_process = subprocess.Popen(args, stdin=subprocess.PIPE, stderr=subprocess.DEVNULL, stdout=subprocess.DEVNULL)
-        else:
-            self.ffmpg_process = subprocess.Popen(args, stdin=subprocess.PIPE)
-        self.init_done = True
-        self.shape_hw = tuple(self.shape_hw)
-        print(f"Initialization done. Movie shape: {self.shape_hw}")
-
-    def write_frame(self, out_frame: np.ndarray):
-        r"""
-        Function to dump a numpy array as frame of a movie.
-        Args:
-            out_frame: np.ndarray
-                Numpy array, in np.uint8 format. Convert with np.astype(x, np.uint8).
-                Dim 0: y
-                Dim 1: x
-                Dim 2: RGB
-        """
-        assert out_frame.dtype == np.uint8, "Convert to np.uint8 before"
-        assert len(out_frame.shape) == 3, "out_frame needs to be three dimensional, Y X C"
-        assert out_frame.shape[2] == 3, f"need three color channels, but you provided {out_frame.shape[2]}."
-
-        if not self.init_done:
-            self.shape_hw = out_frame.shape
-            self.initialize()
-
-        assert self.shape_hw == out_frame.shape, f"You cannot change the image size after init. Initialized with {self.shape_hw}, out_frame {out_frame.shape}"
-
-        # write frame
-        self.ffmpg_process.stdin.write(
-            out_frame
-            .astype(np.uint8)
-            .tobytes()
-        )
-
-        self.nmb_frames += 1
-
-    def finalize(self):
-        r"""
-        Call this function to finalize the movie. If you forget to call it your movie will be garbage.
-        """
-        if self.nmb_frames == 0:
-            print("You did not write any frames yet! nmb_frames = 0. Cannot save.")
-            return
-        self.ffmpg_process.stdin.close()
-        self.ffmpg_process.wait()
-        duration = int(self.nmb_frames / self.fps)
-        print(f"Movie saved, {duration}s playtime, watch here: \n{self.fp_out}")
-
-
-def concatenate_movies(fp_final: str, list_fp_movies: List[str]):
-    r"""
-    Concatenate multiple movie segments into one long movie, using ffmpeg.
-
-    Parameters
-    ----------
-    fp_final : str
-        Full path of the final movie file. Should end with .mp4
-    list_fp_movies : list[str]
-        List of full paths of movie segments.
-    """
-    assert fp_final[-4] == ".", "fp_final seems to miss file extension: {fp_final}"
-    for fp in list_fp_movies:
-        assert os.path.isfile(fp), f"Input movie does not exist: {fp}"
-        assert os.path.getsize(fp) > 100, f"Input movie seems empty: {fp}"
-
-    if os.path.isfile(fp_final):
-        os.remove(fp_final)
-
-    # make a list for ffmpeg
-    list_concat = []
-    for fp_part in list_fp_movies:
-        list_concat.append(f"""file '{fp_part}'""")
-
-    # save this list
-    fp_list = "tmp_move.txt"
-    with open(fp_list, "w") as fa:
-        for item in list_concat:
-            fa.write("%s\n" % item)
-
-    cmd = f'ffmpeg -f concat -safe 0 -i {fp_list} -c copy {fp_final}'
-    subprocess.call(cmd, shell=True)
-    os.remove(fp_list)
-    if os.path.isfile(fp_final):
-        print(f"concatenate_movies: success! Watch here: {fp_final}")
-        
-    
-def add_sound(fp_final, fp_silentmovie, fp_sound):
-    cmd = f'ffmpeg -i {fp_silentmovie} -i {fp_sound} -c copy -map 0:v:0 -map 1:a:0 {fp_final}'
-    subprocess.call(cmd, shell=True)
-    if os.path.isfile(fp_final):
-        print(f"add_sound: success! Watch here: {fp_final}")
-    
-    
-def add_subtitles_to_video(
-        fp_input: str,
-        fp_output: str,
-        subtitles: list,
-        fontsize: int = 50,
-        font_name: str = "Arial",
-        color: str = 'yellow'
-    ):
-    from moviepy.editor import VideoFileClip, TextClip, CompositeVideoClip
-    r"""
-    Function to add subtitles to a video.
-    
-    Args:
-        fp_input (str): File path of the input video.
-        fp_output (str): File path of the output video with subtitles.
-        subtitles (list): List of dictionaries containing subtitle information 
-            (start, duration, text). Example:
-            subtitles = [
-                {"start": 1, "duration": 3, "text": "hello test"},
-                {"start": 4, "duration": 2, "text": "this works"},
-            ]
-        fontsize (int): Font size of the subtitles.
-        font_name (str): Font name of the subtitles.
-        color (str): Color of the subtitles.
-    """
-    
-    # Check if the input file exists
-    if not os.path.isfile(fp_input):
-        raise FileNotFoundError(f"Input file not found: {fp_input}")
-    
-    # Check the subtitles format and sort them by the start time
-    time_points = []
-    for subtitle in subtitles:
-        if not isinstance(subtitle, dict):
-            raise ValueError("Each subtitle must be a dictionary containing 'start', 'duration' and 'text'.")
-        if not all(key in subtitle for key in ["start", "duration", "text"]):
-            raise ValueError("Each subtitle dictionary must contain 'start', 'duration' and 'text'.")
-        if subtitle['start'] < 0 or subtitle['duration'] <= 0:
-            raise ValueError("'start' should be non-negative and 'duration' should be positive.")
-        time_points.append((subtitle['start'], subtitle['start'] + subtitle['duration']))
-
-    # Check for overlaps
-    time_points.sort()
-    for i in range(1, len(time_points)):
-        if time_points[i][0] < time_points[i - 1][1]:
-            raise ValueError("Subtitle time intervals should not overlap.")
-    
-    # Load the video clip
-    video = VideoFileClip(fp_input)
-    
-    # Create a list to store subtitle clips
-    subtitle_clips = []
-    
-    # Loop through the subtitle information and create TextClip for each
-    for subtitle in subtitles:
-        text_clip = TextClip(subtitle["text"], fontsize=fontsize, color=color, font=font_name)
-        text_clip = text_clip.set_position(('center', 'bottom')).set_start(subtitle["start"]).set_duration(subtitle["duration"])
-        subtitle_clips.append(text_clip)
-    
-    # Overlay the subtitles on the video
-    video = CompositeVideoClip([video] + subtitle_clips)
-    
-    # Write the final clip to a new file
-    video.write_videofile(fp_output)
-
-
-
-class MovieReader():
-    r"""
-    Class to read in a movie.
-    """
-
-    def __init__(self, fp_movie):
-        self.video_player_object = cv2.VideoCapture(fp_movie)
-        self.nmb_frames = int(self.video_player_object.get(cv2.CAP_PROP_FRAME_COUNT))
-        self.fps_movie = int(self.video_player_object.get(cv2.CAP_PROP_FPS))
-        self.shape = [100, 100, 3]
-        self.shape_is_set = False
-
-    def get_next_frame(self):
-        success, image = self.video_player_object.read()
-        if success:
-            if not self.shape_is_set:
-                self.shape_is_set = True
-                self.shape = image.shape
-            return image
-        else:
-            return np.zeros(self.shape)
-
-
-if __name__ == "__main__":
-    fps = 2
-    list_fp_movies = []
-    for k in range(4):
-        fp_movie = f"/tmp/my_random_movie_{k}.mp4"
-        list_fp_movies.append(fp_movie)
-        ms = MovieSaver(fp_movie, fps=fps)
-        for fn in tqdm(range(30)):
-            img = (np.random.rand(512, 1024, 3) * 255).astype(np.uint8)
-            ms.write_frame(img)
-        ms.finalize()
-
-    fp_final = "/tmp/my_concatenated_movie.mp4"
-    concatenate_movies(fp_final, list_fp_movies)

requirements.txt

@@ -1,6 +1,6 @@
 lpips==0.1.4
 opencv-python
-ffmpeg-python
 diffusers==0.25.0
 transformers
 pytest
+accelerate

setup.py

@@ -6,14 +6,14 @@ with open('requirements.txt') as f:
 
 setup(
     name='latentblending',
-    version='0.2',
+    version='0.3',
     url='https://github.com/lunarring/latentblending',
     description='Butter-smooth video transitions',
     long_description=open('README.md').read(),
-    install_requires=required,
-    dependency_links=[
-        'git+https://github.com/lunarring/lunar_tools#egg=lunar_tools'
-    ],
+    install_requires=[
+        'lunar_tools @ git+https://github.com/lunarring/lunar_tools.git#egg=lunar_tools'
+    ] + required,
     include_package_data=False,
 )
 
+