Spaces:
Sleeping
Sleeping
Update utils/keyframe_utils.py
Browse files- utils/keyframe_utils.py +5 -47
utils/keyframe_utils.py
CHANGED
|
@@ -1,12 +1,12 @@
|
|
| 1 |
-
import openai
|
| 2 |
import os
|
| 3 |
import json
|
| 4 |
from pathlib import Path
|
| 5 |
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
|
| 6 |
import torch
|
| 7 |
from PIL import Image
|
|
|
|
| 8 |
|
| 9 |
-
|
| 10 |
|
| 11 |
# Global story context
|
| 12 |
story_context_cn = "《博物馆的全能ACE》是一部拟人化博物馆文物与AI讲解助手互动的短片,讲述太阳人石刻在闭馆后的博物馆中,遇到了新来的AI助手博小翼,两者展开对话,AI展示了自己的多模态讲解能力与文化知识,最终被文物们认可,并一起展开智慧导览服务的故事。该片融合了文物拟人化、夜间博物馆奇妙氛围、科技感界面与中国地方文化元素,风格活泼、具未来感。"
|
|
@@ -40,10 +40,10 @@ def generate_keyframe_prompt(segment):
|
|
| 40 |
speaker = segment.get("speaker", "")
|
| 41 |
narration = segment.get("narration", "")
|
| 42 |
|
| 43 |
-
input_prompt = f"你是一个擅长视觉脚本设计的AI,请基于以下故事整体背景与分镜内容,帮我生成一个适合用于Stable Diffusion图像生成的英文提示词(image prompt
|
| 44 |
|
| 45 |
try:
|
| 46 |
-
response =
|
| 47 |
model="gpt-4o",
|
| 48 |
messages=[
|
| 49 |
{"role": "system", "content": "You are an expert visual prompt designer for image generation."},
|
|
@@ -51,7 +51,7 @@ def generate_keyframe_prompt(segment):
|
|
| 51 |
],
|
| 52 |
temperature=0.7
|
| 53 |
)
|
| 54 |
-
output_text = response
|
| 55 |
if "Negative prompt:" in output_text:
|
| 56 |
prompt, negative = output_text.split("Negative prompt:", 1)
|
| 57 |
else:
|
|
@@ -72,45 +72,3 @@ def generate_keyframe_prompt(segment):
|
|
| 72 |
"prompt": description,
|
| 73 |
"negative_prompt": ""
|
| 74 |
}
|
| 75 |
-
|
| 76 |
-
def generate_all_keyframe_images(script_data, output_dir="keyframes"):
|
| 77 |
-
os.makedirs(output_dir, exist_ok=True)
|
| 78 |
-
keyframe_outputs = []
|
| 79 |
-
|
| 80 |
-
for segment in script_data:
|
| 81 |
-
sd_prompts = generate_keyframe_prompt(segment)
|
| 82 |
-
prompt = sd_prompts["prompt"]
|
| 83 |
-
negative_prompt = sd_prompts["negative_prompt"]
|
| 84 |
-
segment_id = segment.get("segment_id")
|
| 85 |
-
|
| 86 |
-
description = segment.get("description", "")
|
| 87 |
-
use_reference = any(name in description for name in ASSET_IMAGES)
|
| 88 |
-
|
| 89 |
-
if use_reference:
|
| 90 |
-
ref_key = next(k for k in ASSET_IMAGES if k in description)
|
| 91 |
-
init_image = Image.open(ASSET_IMAGES[ref_key]).convert("RGB").resize((512, 512))
|
| 92 |
-
|
| 93 |
-
frame_images = []
|
| 94 |
-
for i in range(3):
|
| 95 |
-
if use_reference:
|
| 96 |
-
image = pipe_img2img(prompt=prompt, image=init_image, negative_prompt=negative_prompt, strength=0.6, guidance_scale=7.5).images[0]
|
| 97 |
-
else:
|
| 98 |
-
image = pipe_txt2img(prompt, negative_prompt=negative_prompt, num_inference_steps=20, guidance_scale=7.5, height=256, width=256).images[0]
|
| 99 |
-
|
| 100 |
-
image_path = os.path.join(output_dir, f"segment_{segment_id}_v{i+1}.png")
|
| 101 |
-
image.save(image_path)
|
| 102 |
-
frame_images.append(image_path)
|
| 103 |
-
|
| 104 |
-
keyframe_outputs.append({
|
| 105 |
-
"segment_id": segment_id,
|
| 106 |
-
"prompt": prompt,
|
| 107 |
-
"negative_prompt": negative_prompt,
|
| 108 |
-
"frame_images": frame_images
|
| 109 |
-
})
|
| 110 |
-
|
| 111 |
-
print(f"✓ Generated 3 images for Segment {segment_id} ({'img2img' if use_reference else 'txt2img'})")
|
| 112 |
-
|
| 113 |
-
with open("all_prompts_output.json", "w", encoding="utf-8") as f:
|
| 114 |
-
json.dump(keyframe_outputs, f, ensure_ascii=False, indent=2)
|
| 115 |
-
|
| 116 |
-
return keyframe_outputs
|
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
import json
|
| 3 |
from pathlib import Path
|
| 4 |
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
|
| 5 |
import torch
|
| 6 |
from PIL import Image
|
| 7 |
+
from openai import OpenAI
|
| 8 |
|
| 9 |
+
client = OpenAI()
|
| 10 |
|
| 11 |
# Global story context
|
| 12 |
story_context_cn = "《博物馆的全能ACE》是一部拟人化博物馆文物与AI讲解助手互动的短片,讲述太阳人石刻在闭馆后的博物馆中,遇到了新来的AI助手博小翼,两者展开对话,AI展示了自己的多模态讲解能力与文化知识,最终被文物们认可,并一起展开智慧导览服务的故事。该片融合了文物拟人化、夜间博物馆奇妙氛围、科技感界面与中国地方文化元素,风格活泼、具未来感。"
|
|
|
|
| 40 |
speaker = segment.get("speaker", "")
|
| 41 |
narration = segment.get("narration", "")
|
| 42 |
|
| 43 |
+
input_prompt = f"你是一个擅长视觉脚本设计的AI,请基于以下故事整体背景与分镜内容,帮我生成一个适合用于Stable Diffusion图像生成的英文提示词(image prompt),用于生成低分辨率草图风格的关键帧。请注意突出主要角色、镜头氛围、光影、构图、动作,避免复杂背景和细节。提示词长度不应超过80词,以防止超出Stable Diffusion的token限制。\n\n【整体故事背景】:\n{story_context_cn}\n\n【当前分镜描述】:\n{description}\n【角色】:{speaker}\n【台词或画外音】:{narration}\n\n{REFERENCE_CONTEXT}\n\n请用英文输出一个简洁但具体的prompt,风格偏草图、线稿、卡通、简洁构图,并指出一个negative prompt。"
|
| 44 |
|
| 45 |
try:
|
| 46 |
+
response = client.chat.completions.create(
|
| 47 |
model="gpt-4o",
|
| 48 |
messages=[
|
| 49 |
{"role": "system", "content": "You are an expert visual prompt designer for image generation."},
|
|
|
|
| 51 |
],
|
| 52 |
temperature=0.7
|
| 53 |
)
|
| 54 |
+
output_text = response.choices[0].message.content
|
| 55 |
if "Negative prompt:" in output_text:
|
| 56 |
prompt, negative = output_text.split("Negative prompt:", 1)
|
| 57 |
else:
|
|
|
|
| 72 |
"prompt": description,
|
| 73 |
"negative_prompt": ""
|
| 74 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|