Spaces:
Build error
Build error
File size: 7,001 Bytes
43b5157 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 |
import torch
from enum import Enum
import gc
import numpy as np
import jax.numpy as jnp
import jax
from PIL import Image
from typing import List
from flax.training.common_utils import shard
from flax.jax_utils import replicate
from flax import jax_utils
import einops
from transformers import CLIPTokenizer, CLIPFeatureExtractor, FlaxCLIPTextModel
from diffusers import (
FlaxDDIMScheduler,
FlaxAutoencoderKL,
FlaxUNet2DConditionModel as VanillaFlaxUNet2DConditionModel,
)
from text_to_animation.models.unet_2d_condition_flax import FlaxUNet2DConditionModel
from diffusers import FlaxControlNetModel
from text_to_animation.pipelines.text_to_video_pipeline_flax import (
FlaxTextToVideoPipeline,
)
import utils.utils as utils
import utils.gradio_utils as gradio_utils
import os
on_huggingspace = os.environ.get("SPACE_AUTHOR_NAME") == "PAIR"
unshard = lambda x: einops.rearrange(x, "d b ... -> (d b) ...")
class ModelType(Enum):
Text2Video = 1
ControlNetPose = 2
StableDiffusion = 3
def replicate_devices(array):
return jnp.expand_dims(array, 0).repeat(jax.device_count(), 0)
class ControlAnimationModel:
def __init__(self, dtype, **kwargs):
self.dtype = dtype
self.rng = jax.random.PRNGKey(0)
self.pipe = None
self.model_type = None
self.states = {}
self.model_name = ""
def set_model(
self,
model_id: str,
**kwargs,
):
if hasattr(self, "pipe") and self.pipe is not None:
del self.pipe
self.pipe = None
gc.collect()
controlnet, controlnet_params = FlaxControlNetModel.from_pretrained(
"fusing/stable-diffusion-v1-5-controlnet-openpose",
from_pt=True,
dtype=jnp.float16,
)
scheduler, scheduler_state = FlaxDDIMScheduler.from_pretrained(
model_id, subfolder="scheduler", from_pt=True
)
tokenizer = CLIPTokenizer.from_pretrained(model_id, subfolder="tokenizer")
feature_extractor = CLIPFeatureExtractor.from_pretrained(
model_id, subfolder="feature_extractor"
)
unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(
model_id, subfolder="unet", from_pt=True, dtype=self.dtype
)
unet_vanilla = VanillaFlaxUNet2DConditionModel.from_config(
model_id, subfolder="unet", from_pt=True, dtype=self.dtype
)
vae, vae_params = FlaxAutoencoderKL.from_pretrained(
model_id, subfolder="vae", from_pt=True, dtype=self.dtype
)
text_encoder = FlaxCLIPTextModel.from_pretrained(
model_id, subfolder="text_encoder", from_pt=True, dtype=self.dtype
)
self.pipe = FlaxTextToVideoPipeline(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
unet_vanilla=unet_vanilla,
controlnet=controlnet,
scheduler=scheduler,
safety_checker=None,
feature_extractor=feature_extractor,
)
self.params = {
"unet": unet_params,
"vae": vae_params,
"scheduler": scheduler_state,
"controlnet": controlnet_params,
"text_encoder": text_encoder.params,
}
self.p_params = jax_utils.replicate(self.params)
self.model_name = model_id
def generate_initial_frames(
self,
prompt: str,
video_path: str,
n_prompt: str = "",
num_imgs: int = 4,
resolution: int = 512,
model_id: str = "runwayml/stable-diffusion-v1-5",
) -> List[Image.Image]:
self.set_model(model_id=model_id)
video_path = gradio_utils.motion_to_video_path(video_path)
added_prompt = "high quality, best quality, HD, clay stop-motion, claymation, HQ, masterpiece, art, smooth"
prompts = added_prompt + ", " + prompt
added_n_prompt = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly"
negative_prompts = added_n_prompt + ", " + n_prompt
video, fps = utils.prepare_video(
video_path, resolution, None, self.dtype, False, output_fps=4
)
control = utils.pre_process_pose(video, apply_pose_detect=False)
seeds = [seed for seed in jax.random.randint(self.rng, [num_imgs], 0, 65536)]
prngs = [jax.random.PRNGKey(seed) for seed in seeds]
print(seeds)
images = self.pipe.generate_starting_frames(
params=self.p_params,
prngs=prngs,
controlnet_image=control,
prompt=prompts,
neg_prompt=negative_prompts,
)
images = [np.array(images[i]) for i in range(images.shape[0])]
return images
def generate_video_from_frame(self, controlnet_video, prompt, seed, neg_prompt=""):
# generate a video using the seed provided
prng_seed = jax.random.PRNGKey(seed)
len_vid = controlnet_video.shape[0]
# print(f"Generating video from prompt {'<aardman> style '+ prompt}, with {controlnet_video.shape[0]} frames and prng seed {seed}")
added_prompt = "high quality, best quality, HD, clay stop-motion, claymation, HQ, masterpiece, art, smooth"
prompts = added_prompt + ", " + prompt
added_n_prompt = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly"
negative_prompts = added_n_prompt + ", " + neg_prompt
# prompt_ids = self.pipe.prepare_text_inputs(["aardman style "+ prompt]*len_vid)
# n_prompt_ids = self.pipe.prepare_text_inputs([neg_prompt]*len_vid)
prompt_ids = self.pipe.prepare_text_inputs([prompts] * len_vid)
n_prompt_ids = self.pipe.prepare_text_inputs([negative_prompts] * len_vid)
prng = replicate_devices(
prng_seed
) # jax.random.split(prng, jax.device_count())
image = replicate_devices(controlnet_video)
prompt_ids = replicate_devices(prompt_ids)
n_prompt_ids = replicate_devices(n_prompt_ids)
motion_field_strength_x = replicate_devices(jnp.array(3))
motion_field_strength_y = replicate_devices(jnp.array(4))
smooth_bg_strength = replicate_devices(jnp.array(0.8))
vid = (
self.pipe(
image=image,
prompt_ids=prompt_ids,
neg_prompt_ids=n_prompt_ids,
params=self.p_params,
prng_seed=prng,
jit=True,
smooth_bg_strength=smooth_bg_strength,
motion_field_strength_x=motion_field_strength_x,
motion_field_strength_y=motion_field_strength_y,
).images
)[0]
return utils.create_gif(np.array(vid), 4, path=None, watermark=None)
|