iceDonkey's picture
Duplicate from nvidia/Cosmos3-Super-Text2Image
03f0acb
"""Network clients for standalone agentic text-to-image upsampling."""
from __future__ import annotations
import base64
import io
import json
import os
import time
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import requests
from PIL import Image
from requests.adapters import HTTPAdapter
from urllib3.util.retry import Retry
from agentic_upsampling.constants import (
DEFAULT_ASPECT_RATIO,
DEFAULT_CRITIC_ENDPOINT_URL,
DEFAULT_CRITIC_MODEL,
DEFAULT_GENERATION_AUTH_KEY_ENV,
DEFAULT_GENERATION_EXTRA_ARGS,
DEFAULT_GENERATION_MODEL,
DEFAULT_FLOW_SHIFT,
DEFAULT_GUIDANCE,
DEFAULT_IMAGE_SIZE,
DEFAULT_JPEG_QUALITY,
DEFAULT_LLM_EXTRA_BODY,
DEFAULT_NUM_STEPS,
DEFAULT_OPENAI_API_KEY_ENV,
DEFAULT_RESOLUTION,
DEFAULT_REWRITER_ENDPOINT_URL,
DEFAULT_REWRITER_MODEL,
DEFAULT_UPSAMPLER_ENDPOINT_URL,
DEFAULT_UPSAMPLER_MODEL,
)
from agentic_upsampling.data import PromptItem, validate_t2i_json
from agentic_upsampling.io_utils import compact_json, write_json_atomic
from agentic_upsampling.prompt_upsampler import (
JSON_ENSURE_ASCII,
SYSTEM_MESSAGE,
ChatClientConfig,
OpenAIChatClient,
Text2ImagePromptUpsampler,
extract_json_object,
)
from agentic_upsampling.rubric import (
all_category_check_text,
analysis_json_text,
build_judge_prompt,
compact_analysis_for_rewrite,
parse_analysis_response,
)
CONNECT_TIMEOUT_S = 60
SUBMIT_READ_TIMEOUT_S = 240
IMAGE_GENERATION_READ_TIMEOUT_S = 600
REWRITER_APPLICATION_GUIDANCE = all_category_check_text()
@dataclass(frozen=True, slots=True)
class GenerationOutput:
"""Output from one image generation request."""
image_path: Path
meta_path: Path
meta: dict[str, Any]
def read_api_token(api_key_env: str, api_key_file: Path | None = None) -> str:
"""Resolve an API token from an environment variable or explicit file."""
token = os.environ.get(api_key_env, "").strip()
if token:
return token
if api_key_file is not None and api_key_file.exists():
token = api_key_file.read_text(encoding="utf-8").strip()
if token:
return token
raise RuntimeError(f"Missing API key. Export {api_key_env} or pass the matching --*-api-key-file flag.")
def read_optional_generation_auth_key(auth_key: str, api_key_env: str = DEFAULT_GENERATION_AUTH_KEY_ENV) -> str:
"""Resolve the optional generation endpoint auth key."""
return auth_key.strip() or os.environ.get(api_key_env, "").strip()
def normalize_generation_endpoint(endpoint: str) -> str:
"""Normalize the vLLM-Omni endpoint root without the /v1 suffix."""
normalized = endpoint.strip().rstrip("/")
if not normalized:
raise ValueError("generation endpoint cannot be empty.")
if not normalized.startswith(("http://", "https://")):
normalized = f"https://{normalized}"
if normalized.endswith("/v1/images/generations"):
normalized = normalized[: -len("/v1/images/generations")]
elif normalized.endswith("/v1"):
normalized = normalized[: -len("/v1")]
return normalized.rstrip("/")
def make_session(pool_size: int = 4) -> requests.Session:
"""Create a retrying HTTP session."""
session = requests.Session()
retry = Retry(
total=2,
connect=2,
read=0,
status=2,
status_forcelist=(429, 500, 502, 503, 504),
allowed_methods=frozenset({"GET", "POST"}),
backoff_factor=0.5,
raise_on_status=False,
)
adapter = HTTPAdapter(pool_connections=pool_size, pool_maxsize=pool_size, max_retries=retry, pool_block=False)
session.mount("https://", adapter)
session.mount("http://", adapter)
return session
def image_path_to_data_url(path: Path, *, jpeg_quality: int | None = DEFAULT_JPEG_QUALITY) -> str:
"""Encode a local image file as a data URL, optionally transcoding to JPEG."""
if jpeg_quality is None:
encoded = base64.b64encode(path.read_bytes()).decode("ascii")
return f"data:image/png;base64,{encoded}"
with Image.open(path) as image:
if image.mode not in ("RGB", "L"):
image = image.convert("RGB")
buf = io.BytesIO()
image.save(buf, format="JPEG", quality=jpeg_quality, optimize=True)
encoded = base64.b64encode(buf.getvalue()).decode("ascii")
return f"data:image/jpeg;base64,{encoded}"
class PromptRewriterClient:
"""GPT-based T2I JSON prompt upsampler and iterative rewriter."""
upsampler: Text2ImagePromptUpsampler
rewrite_client: OpenAIChatClient
resolution: str
aspect_ratio: str
def __init__(
self,
*,
api_token: str,
upsampler_endpoint_url: str = DEFAULT_UPSAMPLER_ENDPOINT_URL,
upsampler_model: str = DEFAULT_UPSAMPLER_MODEL,
rewriter_endpoint_url: str = DEFAULT_REWRITER_ENDPOINT_URL,
rewriter_model: str = DEFAULT_REWRITER_MODEL,
extra_body: dict[str, Any] | None = None,
resolution: str = DEFAULT_RESOLUTION,
aspect_ratio: str = DEFAULT_ASPECT_RATIO,
) -> None:
resolved_extra_body = DEFAULT_LLM_EXTRA_BODY if extra_body is None else extra_body
self.upsampler = Text2ImagePromptUpsampler.from_defaults(
api_token=api_token,
endpoint_url=upsampler_endpoint_url,
model=upsampler_model,
extra_body=resolved_extra_body,
)
self.rewrite_client = OpenAIChatClient(
ChatClientConfig(
endpoint_url=rewriter_endpoint_url,
model=rewriter_model,
api_token=api_token,
extra_body=resolved_extra_body,
max_tokens=8192,
max_retries=3,
)
)
self.resolution = resolution
self.aspect_ratio = aspect_ratio
def initial_prompt(self, item: PromptItem) -> dict[str, Any]:
"""Create the initial dense structured prompt for a user prompt."""
return self.upsampler.upsample(
item.prompt,
prompt_id=item.prompt_id,
resolution=self.resolution,
aspect_ratio=self.aspect_ratio,
)
def rewrite_prompt_pair(
self,
item: PromptItem,
previous_prompt: dict[str, Any],
previous_negative_prompt: str,
previous_analysis: dict[str, Any],
history: list[dict[str, Any]],
) -> tuple[dict[str, Any], str]:
"""Jointly rewrite the positive JSON prompt and generator-side negative prompt."""
schema_keys = list(previous_prompt.keys())
messages = [
{
"role": "system",
"content": (
"You are a precise text-to-image prompt engineer. Return valid JSON only, no markdown. "
"Jointly coordinate the positive structured prompt and generator-side negative prompt so they do not contradict each other."
),
},
{
"role": "user",
"content": self._joint_rewrite_user_prompt(
item=item,
previous_prompt=previous_prompt,
previous_negative_prompt=previous_negative_prompt,
previous_analysis=previous_analysis,
history=history,
schema_keys=schema_keys,
),
},
]
last_exc: Exception | None = None
for attempt in range(1, 4):
try:
raw = self.rewrite_client.complete(messages, response_format_json=True)
return self._parse_joint_rewrite_response(raw, item.prompt_id)
except Exception as exc:
last_exc = exc
if attempt < 3:
time.sleep(min(20.0, 2.0 * attempt))
raise RuntimeError(f"Joint prompt rewrite failed after 3 attempts for prompt {item.prompt_id}.") from last_exc
@staticmethod
def _parse_joint_rewrite_response(raw: str, prompt_id: str) -> tuple[dict[str, Any], str]:
data = extract_json_object(raw)
positive_prompt = data.get("positive_prompt")
if not isinstance(positive_prompt, dict):
raise ValueError(f"Joint rewrite returned missing or non-object positive_prompt for prompt {prompt_id}.")
validate_t2i_json(positive_prompt, prompt_id)
negative_prompt = data.get("negative_prompt", "")
if not isinstance(negative_prompt, str):
raise ValueError(f"Joint rewrite returned non-string negative_prompt for prompt {prompt_id}.")
return positive_prompt, " ".join(negative_prompt.split())
@staticmethod
def _joint_rewrite_user_prompt(
*,
item: PromptItem,
previous_prompt: dict[str, Any],
previous_negative_prompt: str,
previous_analysis: dict[str, Any],
history: list[dict[str, Any]],
schema_keys: list[str],
) -> str:
sections = [
"Original user prompt:",
item.prompt,
"",
"Application-specific guidance:",
"Apply the following sections as one checklist program. Do not first classify the prompt. Apply each section only when relevant to the original user prompt, previous JSON, or VLM failures.",
REWRITER_APPLICATION_GUIDANCE,
"",
"Previous generated image failed or scored according to this VLM analysis:",
analysis_json_text(compact_analysis_for_rewrite(previous_analysis)),
"",
"Iteration history summary:",
json.dumps(PromptRewriterClient._history_summary(history), ensure_ascii=JSON_ENSURE_ASCII, indent=2),
"",
"Previous positive JSON prompt:",
json.dumps(previous_prompt, ensure_ascii=JSON_ENSURE_ASCII, indent=2),
"",
"Previous negative prompt:",
previous_negative_prompt or "",
"",
"Joint rewrite task:",
'Return a JSON object with exactly two top-level keys: "positive_prompt" and "negative_prompt".',
'"positive_prompt" must be a complete JSON object with exactly these top-level keys, preserving their names and types:',
json.dumps(schema_keys, ensure_ascii=JSON_ENSURE_ASCII),
"",
'"positive_prompt" must keep the previous "resolution" and "aspect_ratio".',
'"negative_prompt" must be a concise generator-side negative prompt string.',
"Coordinate both fields: strengthen required positive constraints while using the negative prompt only to suppress concrete wrong alternatives or artifacts.",
"Do not put positive instructions in negative_prompt. Do not negate content required by the original user prompt.",
"For exact counts, grids, text, geometry, or anatomy, explicitly block wrong alternatives when useful.",
'The positive "comprehensive_t2i_caption" should be direct generation guidance, not an explanation of this rewrite process.',
]
return "\n".join(sections)
@staticmethod
def _history_summary(history: list[dict[str, Any]]) -> list[dict[str, Any]]:
return [
{
"iteration": item.get("iteration"),
"overall_score": item.get("analysis", {}).get("overall_score"),
"prompt_adherence_score": item.get("analysis", {}).get("prompt_adherence_score"),
"category_score": item.get("analysis", {}).get("category_score"),
"threshold_cleared": item.get("analysis", {}).get("threshold_cleared"),
}
for item in history
]
class ImageGenerationClient:
"""Client for a vLLM-Omni /v1/images/generations text-to-image endpoint."""
endpoint: str
auth_key: str
model: str
session: requests.Session
size: str
num_steps: int
guidance: float
flow_shift: float
extra_args: dict[str, Any]
def __init__(
self,
*,
endpoint: str,
auth_key: str = "",
model: str = DEFAULT_GENERATION_MODEL,
size: str = DEFAULT_IMAGE_SIZE,
num_steps: int = DEFAULT_NUM_STEPS,
guidance: float = DEFAULT_GUIDANCE,
flow_shift: float = DEFAULT_FLOW_SHIFT,
extra_args: dict[str, Any] | None = None,
session: requests.Session | None = None,
) -> None:
self.endpoint = normalize_generation_endpoint(endpoint)
self.auth_key = auth_key
self.model = model
self.session = session or make_session()
self.size = size
self.num_steps = num_steps
self.guidance = guidance
self.flow_shift = flow_shift
self.extra_args = dict(DEFAULT_GENERATION_EXTRA_ARGS if extra_args is None else extra_args)
def build_payload(
self,
prompt_json: dict[str, Any],
prompt_id: str,
seed: int | None = None,
negative_prompt: str = "",
) -> dict[str, Any]:
"""Build the vLLM-Omni image generation request payload."""
del prompt_id
payload: dict[str, Any] = {
"model": self.model,
"prompt": compact_json(prompt_json, ensure_ascii=JSON_ENSURE_ASCII),
"size": self.size,
"n": 1,
"response_format": "b64_json",
"negative_prompt": negative_prompt.strip(),
"num_inference_steps": self.num_steps,
"guidance_scale": self.guidance,
"flow_shift": self.flow_shift,
"extra_args": dict(self.extra_args),
}
if seed is not None:
payload["seed"] = int(seed)
return payload
def generate(
self,
*,
prompt_json: dict[str, Any],
prompt_id: str,
output_dir: Path,
seed: int | None = None,
negative_prompt: str = "",
jpeg_quality: int = DEFAULT_JPEG_QUALITY,
) -> GenerationOutput:
"""Generate and persist one candidate image."""
payload = self.build_payload(prompt_json, prompt_id, seed, negative_prompt=negative_prompt)
response_json = self._generate_image(payload)
image_bytes = self._decode_image_response(response_json)
image_path = output_dir / "image.jpg"
image_info = self._save_jpeg(image_bytes, image_path, jpeg_quality)
meta = {
"prompt_id": prompt_id,
"status": "completed",
"endpoint": self.endpoint,
"image_generation_url": self._image_generation_url(),
"payload": payload,
"response": self._response_without_image_bytes(response_json),
"output_image_path": str(image_path),
"image_info": image_info,
}
meta_path = output_dir / "generation_meta.json"
write_json_atomic(meta_path, meta, ensure_ascii=JSON_ENSURE_ASCII)
return GenerationOutput(image_path=image_path, meta_path=meta_path, meta=meta)
def _generate_image(self, payload: dict[str, Any]) -> dict[str, Any]:
last_exc: Exception | None = None
for attempt in range(1, 4):
try:
return self._request_json(
"POST",
self._image_generation_url(),
json=payload,
headers=self._auth_headers(),
timeout=(CONNECT_TIMEOUT_S, IMAGE_GENERATION_READ_TIMEOUT_S),
)
except Exception as exc:
last_exc = exc
if attempt < 3:
time.sleep(min(20.0, 2.0 * attempt))
raise RuntimeError(f"/v1/images/generations failed after retries: {last_exc}") from last_exc
def _image_generation_url(self) -> str:
return f"{self.endpoint}/v1/images/generations"
def _auth_headers(self) -> dict[str, str] | None:
token = self.auth_key.strip()
if not token:
return None
if token.lower().startswith("bearer "):
return {"Authorization": token}
return {"Authorization": f"Bearer {token}"}
def _request_json(self, method: str, url: str, **kwargs: Any) -> dict[str, Any]:
timeout = kwargs.pop("timeout", (CONNECT_TIMEOUT_S, IMAGE_GENERATION_READ_TIMEOUT_S))
response = self.session.request(method, url, timeout=timeout, **kwargs)
if not response.ok:
raise RuntimeError(f"{method} {url} HTTP {response.status_code}: {response.text[:1000]}")
parsed = response.json()
if not isinstance(parsed, dict):
raise RuntimeError(f"{method} {url} returned non-object JSON: {parsed!r}")
return parsed
@staticmethod
def _decode_image_response(response_json: dict[str, Any]) -> bytes:
data = response_json.get("data")
if not isinstance(data, list) or not data or not isinstance(data[0], dict):
raise RuntimeError(f"Image generation response has no data[0] object: {response_json}")
first_image = data[0]
b64_image = first_image.get("b64_json")
if not isinstance(b64_image, str) or not b64_image.strip():
image_url = first_image.get("url")
if isinstance(image_url, str) and image_url.startswith("data:image") and "," in image_url:
b64_image = image_url.split(",", 1)[1]
else:
raise RuntimeError(f"Image generation response has no b64_json image: {response_json}")
try:
return base64.b64decode(b64_image, validate=True)
except ValueError:
return base64.b64decode(b64_image)
@staticmethod
def _response_without_image_bytes(response_json: dict[str, Any]) -> dict[str, Any]:
redacted = json.loads(json.dumps(response_json))
data = redacted.get("data")
if isinstance(data, list):
for item in data:
if isinstance(item, dict) and isinstance(item.get("b64_json"), str):
item["b64_json"] = f"<base64 image omitted: {len(item['b64_json'])} chars>"
if isinstance(item, dict) and isinstance(item.get("url"), str) and item["url"].startswith("data:image"):
item["url"] = f"<data image omitted: {len(item['url'])} chars>"
return redacted
@staticmethod
def _save_jpeg(image_bytes: bytes, output_path: Path, quality: int) -> dict[str, Any]:
output_path.parent.mkdir(parents=True, exist_ok=True)
tmp = output_path.with_suffix(output_path.suffix + ".tmp")
with Image.open(io.BytesIO(image_bytes)) as image:
source_format = image.format
rgb = image.convert("RGB")
width, height = rgb.size
rgb.save(tmp, format="JPEG", quality=quality, optimize=True)
tmp.replace(output_path)
return {"source_image_format": source_format, "saved_format": "JPEG", "width": width, "height": height}
class VLMQualityJudge:
"""Gemini critic for generated images through an OpenAI-compatible endpoint."""
chat_client: OpenAIChatClient
image_jpeg_quality: int | None
def __init__(
self,
*,
api_token: str,
endpoint_url: str = DEFAULT_CRITIC_ENDPOINT_URL,
model: str = DEFAULT_CRITIC_MODEL,
max_tokens: int = 8192,
image_jpeg_quality: int | None = DEFAULT_JPEG_QUALITY,
) -> None:
self.chat_client = OpenAIChatClient(
ChatClientConfig(
endpoint_url=endpoint_url,
model=model,
api_token=api_token,
max_tokens=max_tokens,
max_retries=3,
)
)
self.image_jpeg_quality = image_jpeg_quality
def score_image(
self,
*,
item: PromptItem,
image_path: Path,
) -> dict[str, Any]:
"""Score one image with the non-classifying rubric program."""
messages = [
SYSTEM_MESSAGE,
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {"url": image_path_to_data_url(image_path, jpeg_quality=self.image_jpeg_quality)},
},
{
"type": "text",
"text": build_judge_prompt(item),
},
],
},
]
raw = self.chat_client.complete(messages, response_format_json=True)
analysis = parse_analysis_response(raw)
analysis["raw_response"] = raw
return analysis