Files
AI-Video/pixelle_video/services/quality/character_analyzer.py

324 lines
13 KiB
Python

# Copyright (C) 2025 AIDC-AI
#
# 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.
"""
CharacterAnalyzer - VLM-based character appearance extraction
Analyzes reference images to extract detailed character descriptions
for maintaining visual consistency across video frames.
"""
import base64
import json
import os
import re
from dataclasses import dataclass
from typing import List, Optional
from loguru import logger
from openai import AsyncOpenAI
@dataclass
class CharacterAnalysisResult:
"""Result of character image analysis"""
appearance_description: str = "" # Physical features
clothing_description: str = "" # What they're wearing
distinctive_features: List[str] = None # Unique identifying features
def __post_init__(self):
if self.distinctive_features is None:
self.distinctive_features = []
def to_prompt_description(self) -> str:
"""Generate a prompt-ready character description"""
parts = []
if self.appearance_description:
parts.append(self.appearance_description)
if self.clothing_description:
parts.append(f"wearing {self.clothing_description}")
if self.distinctive_features:
features = ", ".join(self.distinctive_features)
parts.append(f"with {features}")
return ", ".join(parts) if parts else ""
def to_dict(self) -> dict:
return {
"appearance_description": self.appearance_description,
"clothing_description": self.clothing_description,
"distinctive_features": self.distinctive_features,
}
class CharacterAnalyzer:
"""
VLM-based character appearance analyzer
Analyzes reference images to extract detailed character descriptions
that can be injected into image generation prompts.
Example:
>>> analyzer = CharacterAnalyzer()
>>> result = await analyzer.analyze_reference_image("character.png")
>>> print(result.appearance_description)
"young woman with long black hair, round face, fair skin"
>>> print(result.to_prompt_description())
"young woman with long black hair, round face, fair skin, wearing blue hoodie, with round glasses"
"""
def __init__(self):
"""Initialize CharacterAnalyzer"""
pass
async def analyze_reference_image(
self,
image_path: str,
) -> CharacterAnalysisResult:
"""
Analyze a reference image to extract character appearance
Args:
image_path: Path to the reference image
Returns:
CharacterAnalysisResult with extracted descriptions
"""
logger.info(f"Analyzing character reference image: {image_path}")
# Check if file exists
if not os.path.exists(image_path):
logger.warning(f"Image not found: {image_path}")
return CharacterAnalysisResult()
try:
# Read and encode image
with open(image_path, "rb") as f:
image_data = base64.b64encode(f.read()).decode("utf-8")
# Determine image type
ext = os.path.splitext(image_path)[1].lower()
media_type = "image/png" if ext == ".png" else "image/jpeg"
# VLM prompt for character analysis - optimized for storyboard consistency
# Focus on CONSTANT features (identity), exclude VARIABLE features (pose/expression)
analysis_prompt = """Analyze this character/person for VIDEO STORYBOARD consistency.
GOAL: Extract features that should remain CONSISTENT across different video frames.
The output will be injected into image generation prompts for multiple scenes.
Extract ONLY these CONSTANT features:
1. Identity: gender, approximate age group (child/young/middle-aged/elderly)
2. Hair: color, length, style (NOT affected by wind/movement)
3. Face: skin tone, face shape (NOT expressions)
4. Clothing: type and colors (assume same outfit throughout video)
5. Distinctive: glasses, accessories, tattoos, scars, unique marks
DO NOT include:
- Expressions (smile, frown) - changes per scene
- Poses/gestures - changes per scene
- View angle - determined by scene composition
- Lighting/shadows - scene-dependent
- Background elements
Output JSON format (simple strings for direct prompt injection):
{
"identity": "elderly man" or "young woman" etc,
"appearance": "short gray hair, light skin, round face",
"clothing": "brown sweater vest over white shirt, dark trousers",
"distinctive": ["round glasses", "silver watch"]
}
Output ONLY the JSON, no explanation."""
# Build multimodal message
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": analysis_prompt},
{
"type": "image_url",
"image_url": {
"url": f"data:{media_type};base64,{image_data}"
}
}
]
}
]
# Get VLM configuration
# Priority: Environment variables > config.yaml > defaults
from pixelle_video.config import config_manager
# VLM config from config.yaml (now part of PixelleVideoConfig)
vlm_config = config_manager.config.vlm
# Environment variables override config.yaml
vlm_provider = os.getenv("VLM_PROVIDER") or vlm_config.provider or "qwen"
vlm_api_key = os.getenv("VLM_API_KEY") or os.getenv("DASHSCOPE_API_KEY") or vlm_config.api_key
vlm_base_url = os.getenv("VLM_BASE_URL") or vlm_config.base_url
vlm_model = os.getenv("VLM_MODEL") or vlm_config.model
# Configure based on provider
if vlm_provider == "qwen":
# 通义千问 Qwen VL
vlm_base_url = vlm_base_url or "https://dashscope.aliyuncs.com/compatible-mode/v1"
vlm_model = vlm_model or "qwen-vl-plus" # or qwen-vl-max, qwen3-vl-plus
logger.info(f"Using Qwen VL: model={vlm_model}")
elif vlm_provider == "glm":
# 智谱 GLM-4V
from pixelle_video.config import config_manager
llm_config = config_manager.config.llm
vlm_api_key = vlm_api_key or llm_config.api_key
vlm_base_url = vlm_base_url or llm_config.base_url
vlm_model = vlm_model or "glm-4v-flash"
logger.info(f"Using GLM VL: model={vlm_model}")
else: # openai or other
from pixelle_video.config import config_manager
llm_config = config_manager.config.llm
vlm_api_key = vlm_api_key or llm_config.api_key
vlm_base_url = vlm_base_url or llm_config.base_url
vlm_model = vlm_model or llm_config.model
logger.info(f"Using {vlm_provider} VL: model={vlm_model}")
if not vlm_api_key:
logger.error("No VLM API key configured. Set VLM_API_KEY or DASHSCOPE_API_KEY environment variable.")
return CharacterAnalysisResult()
# Create OpenAI-compatible client
client = AsyncOpenAI(
api_key=vlm_api_key,
base_url=vlm_base_url
)
# Call VLM
response = await client.chat.completions.create(
model=vlm_model,
messages=messages,
temperature=0.3,
max_tokens=2000
)
vlm_response = response.choices[0].message.content if response.choices else None
if vlm_response:
logger.debug(f"VLM character analysis response: {vlm_response[:150] if len(vlm_response) > 150 else vlm_response}...")
else:
logger.warning(f"VLM returned empty content. Full response: {response}")
# Parse response
return self._parse_response(vlm_response)
except Exception as e:
logger.error(f"Character analysis failed: {e}")
return CharacterAnalysisResult()
def _parse_response(self, response: str) -> CharacterAnalysisResult:
"""Parse VLM response into CharacterAnalysisResult"""
if not response:
logger.warning("Empty VLM response")
return CharacterAnalysisResult()
# Log full response for debugging
logger.debug(f"Full VLM response:\n{response}")
try:
# Remove markdown code blocks if present
cleaned = response.strip()
if cleaned.startswith("```json"):
cleaned = cleaned[7:]
elif cleaned.startswith("```"):
cleaned = cleaned[3:]
if cleaned.endswith("```"):
cleaned = cleaned[:-3]
cleaned = cleaned.strip()
# Try to extract JSON from response
match = re.search(r'\{[\s\S]*\}', cleaned)
if match:
json_str = match.group()
logger.debug(f"Extracted JSON: {json_str[:200]}...")
data = json.loads(json_str)
else:
logger.warning(f"No JSON found in response, trying direct parse")
data = json.loads(cleaned)
# Handle nested JSON structures - flatten to strings
# New field names: identity, appearance, clothing, distinctive
identity = data.get("identity", "")
appearance = data.get("appearance", "") or data.get("appearance_description", "")
if isinstance(appearance, dict):
# Flatten nested object to descriptive string
parts = []
for key, value in appearance.items():
if isinstance(value, dict):
details = ", ".join(f"{k}: {v}" for k, v in value.items())
parts.append(f"{key} ({details})")
else:
parts.append(f"{key}: {value}")
appearance = "; ".join(parts)
# Combine identity + appearance for full description
if identity and appearance:
full_appearance = f"{identity}, {appearance}"
else:
full_appearance = identity or appearance
clothing = data.get("clothing", "") or data.get("clothing_description", "")
if isinstance(clothing, dict):
# Flatten nested clothing description
parts = []
for person, items in clothing.items():
if isinstance(items, dict):
details = ", ".join(f"{k}: {v}" for k, v in items.items())
parts.append(f"{person} ({details})")
else:
parts.append(f"{person}: {items}")
clothing = "; ".join(parts)
distinctive = data.get("distinctive", []) or data.get("distinctive_features", [])
if not isinstance(distinctive, list):
distinctive = [str(distinctive)]
result = CharacterAnalysisResult(
appearance_description=full_appearance,
clothing_description=clothing,
distinctive_features=distinctive,
)
logger.info(f"Character analysis extracted: {result.appearance_description[:80] if result.appearance_description else 'empty'}...")
return result
except (json.JSONDecodeError, KeyError) as e:
logger.warning(f"Failed to parse VLM response: {e}")
logger.debug(f"Response that failed to parse: {response[:500]}")
# Try to use the raw response as appearance description (fallback)
if response and 20 < len(response) < 500:
# Clean up the response
fallback = response.strip()
if "```" in fallback:
fallback = re.sub(r'```.*?```', '', fallback, flags=re.DOTALL).strip()
if fallback:
logger.info(f"Using raw response as appearance: {fallback[:80]}...")
return CharacterAnalysisResult(
appearance_description=fallback
)
return CharacterAnalysisResult()