import gradio as gr from PIL import Image, PngImagePlugin # Убедимся, что Image из PIL импортирован import io import os import pandas as pd import torch from transformers import pipeline as transformers_pipeline , CLIPImageProcessor import open_clip import re import matplotlib.pyplot as plt import json from collections import defaultdict import numpy as np import logging import time import tempfile # --- ONNX Related Imports and Setup --- try: import onnxruntime except ImportError: print("WARNING: onnxruntime not found. ONNX models will not be available.") onnxruntime = None from huggingface_hub import hf_hub_download try: from imgutils.data import rgb_encode IMGUTILS_AVAILABLE = True print("INFO: imgutils.data.rgb_encode found and will be used for deepghs models.") except ImportError: print("WARNING: imgutils.data.rgb_encode not found. Using a basic fallback for preprocessing deepghs models.") IMGUTILS_AVAILABLE = False def rgb_encode(image: Image.Image, order_='CHW'): img_arr = np.array(image.convert("RGB")) if order_ == 'CHW': img_arr = np.transpose(img_arr, (2, 0, 1)) return img_arr.astype(np.uint8) DEVICE = "cuda" if torch.cuda.is_available() else "cpu" print(f"INFO: PyTorch Device: {DEVICE}") ONNX_EXECUTION_PROVIDER = "CUDAExecutionProvider" if DEVICE == "cuda" and onnxruntime and "CUDAExecutionProvider" in onnxruntime.get_available_providers() else "CPUExecutionProvider" if onnxruntime: print(f"INFO: ONNX Execution Provider: {ONNX_EXECUTION_PROVIDER}") else: print("INFO: ONNX Runtime not available, ONNX models will be skipped.") @torch.no_grad() def _img_preprocess_for_onnx(image: Image.Image, size: tuple = (384, 384), normalize_mean=0.5, normalize_std=0.5): image = image.resize(size, Image.Resampling.BILINEAR) data_uint8 = rgb_encode(image, order_='CHW') data_float01 = data_uint8.astype(np.float32) / 255.0 mean = np.array([normalize_mean] * 3, dtype=np.float32).reshape((3, 1, 1)) std = np.array([normalize_std] * 3, dtype=np.float32).reshape((3, 1, 1)) normalized_data = (data_float01 - mean) / std return normalized_data[None, ...].astype(np.float32) onnx_sessions_cache = {} def get_onnx_session_and_meta(repo_id, model_subfolder, current_log_list): cache_key = f"{repo_id}/{model_subfolder}" if cache_key in onnx_sessions_cache: return onnx_sessions_cache[cache_key] if not onnxruntime: msg = f"ERROR: ONNX Runtime not available for get_onnx_session_and_meta ({cache_key}). Skipping." print(msg); current_log_list.append(msg) onnx_sessions_cache[cache_key] = (None, [], None) return None, [], None try: msg = f"INFO: Loading ONNX model {repo_id}/{model_subfolder}..." print(msg); current_log_list.append(msg) model_path = hf_hub_download(repo_id, filename=f"{model_subfolder}/model.onnx") meta_path = hf_hub_download(repo_id, filename=f"{model_subfolder}/meta.json") options = onnxruntime.SessionOptions() options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL if ONNX_EXECUTION_PROVIDER == "CPUExecutionProvider" and hasattr(os, 'cpu_count'): options.intra_op_num_threads = os.cpu_count() session = onnxruntime.InferenceSession(model_path, options, providers=[ONNX_EXECUTION_PROVIDER]) with open(meta_path, 'r') as f: meta = json.load(f) labels = meta.get('labels', []) msg = f"INFO: ONNX model {cache_key} loaded successfully with provider {ONNX_EXECUTION_PROVIDER}." print(msg); current_log_list.append(msg) onnx_sessions_cache[cache_key] = (session, labels, meta) return session, labels, meta except Exception as e: msg = f"ERROR: Failed to load ONNX model {cache_key}: {e}" print(msg); current_log_list.append(msg) onnx_sessions_cache[cache_key] = (None, [], None) return None, [], None reward_processor, reward_model = None, None print("INFO: THUDM/ImageReward is temporarily disabled due to loading issues.") ANIME_AESTHETIC_REPO = "deepghs/anime_aesthetic"; ANIME_AESTHETIC_SUBFOLDER = "swinv2pv3_v0_448_ls0.2_x" ANIME_AESTHETIC_IMG_SIZE = (448, 448); ANIME_AESTHETIC_LABEL_WEIGHTS = {"normal": 0.0, "slight": 1.0, "moderate": 2.0, "strong": 3.0, "extreme": 4.0} print("INFO: MANIQA (honklers/maniqa-nr) is currently disabled.") clip_model_instance, clip_preprocess, clip_tokenizer = None, None, None try: clip_model_name = 'ViT-L-14'; print(f"INFO: Loading CLIP model {clip_model_name} (laion2b_s32b_b82k)...") clip_model_instance, _, clip_preprocess_val = open_clip.create_model_and_transforms(clip_model_name, pretrained='laion2b_s32b_b82k', device=DEVICE) clip_preprocess = clip_preprocess_val; clip_tokenizer = open_clip.get_tokenizer(clip_model_name) clip_model_instance.eval(); print(f"INFO: CLIP model {clip_model_name} (laion2b_s32b_b82k) loaded successfully.") except Exception as e: print(f"ERROR: Failed to load CLIP model {clip_model_name} (laion2b_s32b_b82k): {e}") sdxl_detector_pipe = None try: print("INFO: Loading Organika/sdxl-detector model...") sdxl_detector_pipe = transformers_pipeline("image-classification", model="Organika/sdxl-detector", device=torch.device(DEVICE).index if DEVICE=="cuda" else -1) print("INFO: Organika/sdxl-detector loaded successfully.") except Exception as e: print(f"ERROR: Failed to load Organika/sdxl-detector: {e}") ANIME_AI_CHECK_REPO = "deepghs/anime_ai_check"; ANIME_AI_CHECK_SUBFOLDER = "caformer_s36_plus_sce" ANIME_AI_CHECK_IMG_SIZE = (384, 384) def extract_sd_parameters(image_pil, filename_for_log, current_log_list): if image_pil is None: return "", "N/A", "N/A", "N/A", {} parameters_str = image_pil.info.get("parameters", "") if not parameters_str: current_log_list.append(f"DEBUG [{filename_for_log}]: No metadata found in image.") return "", "N/A", "N/A", "N/A", {} current_log_list.append(f"DEBUG [{filename_for_log}]: Raw metadata: {parameters_str[:100]}...") prompt, negative_prompt, model_name, model_hash, other_params_dict = "", "N/A", "N/A", "N/A", {} try: neg_prompt_index = parameters_str.find("Negative prompt:") steps_meta_index = parameters_str.find("Steps:") if neg_prompt_index != -1: prompt = parameters_str[:neg_prompt_index].strip() params_part_start_index = steps_meta_index if steps_meta_index != -1 and steps_meta_index > neg_prompt_index else -1 if params_part_start_index != -1: negative_prompt = parameters_str[neg_prompt_index + len("Negative prompt:"):params_part_start_index].strip() params_part = parameters_str[params_part_start_index:] else: end_of_neg = parameters_str.find("\n", neg_prompt_index + len("Negative prompt:")) if end_of_neg == -1: end_of_neg = len(parameters_str) negative_prompt = parameters_str[neg_prompt_index + len("Negative prompt:"):end_of_neg].strip() params_part = parameters_str[end_of_neg:].strip() if end_of_neg < len(parameters_str) else "" elif steps_meta_index != -1: prompt = parameters_str[:steps_meta_index].strip(); params_part = parameters_str[steps_meta_index:] else: prompt = parameters_str.strip(); params_part = "" if params_part: params_list = [p.strip() for p in params_part.split(",")] temp_other_params = {} for param_val_str in params_list: parts = param_val_str.split(':', 1) if len(parts) == 2: key, value = parts[0].strip(), parts[1].strip() temp_other_params[key] = value if key.lower() == "model": model_name = value elif key.lower() == "model hash": model_hash = value for k,v in temp_other_params.items(): if k.lower() not in ["model", "model hash"]: other_params_dict[k] = v if model_name == "N/A" and model_hash != "N/A": model_name = f"hash_{model_hash}" if model_name == "N/A" and "Checkpoint" in other_params_dict: model_name = other_params_dict["Checkpoint"] if model_name == "N/A" and "model" in other_params_dict: model_name = other_params_dict["model"] current_log_list.append(f"DEBUG [{filename_for_log}]: Parsed Prompt: {prompt[:50]}... | Model: {model_name}") except Exception as e: current_log_list.append(f"ERROR [{filename_for_log}]: Failed to parse metadata: {e}") return prompt, negative_prompt, model_name, model_hash, other_params_dict @torch.no_grad() def get_image_reward(image_pil, filename_for_log, current_log_list): return "N/A (Disabled)" def get_anime_aesthetic_score_deepghs(image_pil, filename_for_log, current_log_list): session, labels, meta = get_onnx_session_and_meta(ANIME_AESTHETIC_REPO, ANIME_AESTHETIC_SUBFOLDER, current_log_list) if not session or not labels: current_log_list.append(f"INFO [{filename_for_log}]: AnimeAesthetic ONNX model not loaded, skipping."); return "N/A" t_start = time.time(); current_log_list.append(f"DEBUG [{filename_for_log}]: Starting AnimeAesthetic (ONNX) score...") try: input_data = _img_preprocess_for_onnx(image_pil.copy(), size=ANIME_AESTHETIC_IMG_SIZE) input_name = session.get_inputs()[0].name; output_name = session.get_outputs()[0].name onnx_output, = session.run([output_name], {input_name: input_data}) scores = onnx_output[0]; exp_scores = np.exp(scores - np.max(scores)); probabilities = exp_scores / np.sum(exp_scores) weighted_score = sum(probabilities[i] * ANIME_AESTHETIC_LABEL_WEIGHTS.get(label, 0.0) for i, label in enumerate(labels)) score = round(weighted_score, 4); t_end = time.time() current_log_list.append(f"DEBUG [{filename_for_log}]: AnimeAesthetic (ONNX) score: {score} (took {t_end - t_start:.2f}s)"); return score except Exception as e: current_log_list.append(f"ERROR [{filename_for_log}]: AnimeAesthetic (ONNX) scoring failed: {e}"); return "Error" @torch.no_grad() def get_maniqa_score(image_pil, filename_for_log, current_log_list): current_log_list.append(f"INFO [{filename_for_log}]: MANIQA is disabled."); return "N/A (Disabled)" @torch.no_grad() def calculate_clip_score_value(image_pil, prompt_text, filename_for_log, current_log_list): if not clip_model_instance or not clip_preprocess or not clip_tokenizer: current_log_list.append(f"INFO [{filename_for_log}]: CLIP model not loaded, skipping CLIPScore."); return "N/A" if not prompt_text or prompt_text == "N/A": current_log_list.append(f"INFO [{filename_for_log}]: Empty prompt, skipping CLIPScore."); return "N/A (Empty Prompt)" t_start = time.time(); current_log_list.append(f"DEBUG [{filename_for_log}]: Starting CLIPScore (PyTorch Device: {DEVICE})...") try: image_input = clip_preprocess(image_pil).unsqueeze(0).to(DEVICE) text_for_tokenizer = str(prompt_text); text_input = clip_tokenizer([text_for_tokenizer]).to(DEVICE) image_features = clip_model_instance.encode_image(image_input); text_features = clip_model_instance.encode_text(text_input) image_features_norm = image_features / image_features.norm(p=2, dim=-1, keepdim=True) text_features_norm = text_features / text_features.norm(p=2, dim=-1, keepdim=True) score_val = (text_features_norm @ image_features_norm.T).squeeze().item() * 100.0 score = round(score_val, 2); t_end = time.time() current_log_list.append(f"DEBUG [{filename_for_log}]: CLIPScore: {score} (took {t_end - t_start:.2f}s)"); return score except Exception as e: current_log_list.append(f"ERROR [{filename_for_log}]: CLIPScore calculation failed: {e}"); return "Error" @torch.no_grad() def get_sdxl_detection_score(image_pil, filename_for_log, current_log_list): if not sdxl_detector_pipe: current_log_list.append(f"INFO [{filename_for_log}]: SDXL_Detector model not loaded, skipping."); return "N/A" t_start = time.time(); current_log_list.append(f"DEBUG [{filename_for_log}]: Starting SDXL_Detector score (Device: {sdxl_detector_pipe.device})...") try: result = sdxl_detector_pipe(image_pil.copy()); ai_score_val = 0.0 for item in result: if item['label'].lower() == 'artificial': ai_score_val = item['score']; break score = round(ai_score_val, 4); t_end = time.time() current_log_list.append(f"DEBUG [{filename_for_log}]: SDXL_Detector AI Prob: {score} (took {t_end - t_start:.2f}s)"); return score except Exception as e: current_log_list.append(f"ERROR [{filename_for_log}]: SDXL_Detector scoring failed: {e}"); return "Error" def get_anime_ai_check_score_deepghs(image_pil, filename_for_log, current_log_list): session, labels, meta = get_onnx_session_and_meta(ANIME_AI_CHECK_REPO, ANIME_AI_CHECK_SUBFOLDER, current_log_list) if not session or not labels: current_log_list.append(f"INFO [{filename_for_log}]: AnimeAI_Check ONNX model not loaded, skipping."); return "N/A" t_start = time.time(); current_log_list.append(f"DEBUG [{filename_for_log}]: Starting AnimeAI_Check (ONNX) score...") try: input_data = _img_preprocess_for_onnx(image_pil.copy(), size=ANIME_AI_CHECK_IMG_SIZE) input_name = session.get_inputs()[0].name; output_name = session.get_outputs()[0].name onnx_output, = session.run([output_name], {input_name: input_data}) scores = onnx_output[0]; exp_scores = np.exp(scores - np.max(scores)); probabilities = exp_scores / np.sum(exp_scores) ai_prob_val = 0.0 for i, label in enumerate(labels): if label.lower() == 'ai': ai_prob_val = probabilities[i]; break score = round(ai_prob_val, 4); t_end = time.time() current_log_list.append(f"DEBUG [{filename_for_log}]: AnimeAI_Check (ONNX) AI Prob: {score} (took {t_end - t_start:.2f}s)"); return score except Exception as e: current_log_list.append(f"ERROR [{filename_for_log}]: AnimeAI_Check (ONNX) scoring failed: {e}"); return "Error" def process_images_generator(files, progress=gr.Progress(track_tqdm=True)): if not files: yield (pd.DataFrame(), gr.Image(visible=False), gr.Image(visible=False), gr.File(visible=False), gr.File(visible=False), "Please upload some images.", "No files to process.") return all_results = [] log_accumulator = [f"INFO: Starting processing for {len(files)} images..."] yield (pd.DataFrame(all_results), gr.Image(visible=False), gr.Image(visible=False), gr.File(visible=False), gr.File(visible=False), "Processing...", "\n".join(log_accumulator)) for i, file_obj in enumerate(files): filename_for_log = "Unknown File"; current_img_total_time_start = time.time() try: filename_for_log = os.path.basename(getattr(file_obj, 'name', f"file_{i}_{int(time.time())}")) log_accumulator.append(f"--- Processing image {i+1}/{len(files)}: {filename_for_log} ---") progress( (i + 0.1) / len(files), desc=f"Img {i+1}/{len(files)}: Loading {filename_for_log}") yield (pd.DataFrame(all_results), gr.Image(visible=False), gr.Image(visible=False), gr.File(visible=False), gr.File(visible=False), f"Loading image {i+1}/{len(files)}: {filename_for_log}", "\n".join(log_accumulator)) img = Image.open(getattr(file_obj, 'name', str(file_obj))) if img.mode != "RGB": img = img.convert("RGB") progress( (i + 0.3) / len(files), desc=f"Img {i+1}/{len(files)}: Scoring {filename_for_log}") prompt, neg_prompt, model_n, model_h, other_p = extract_sd_parameters(img, filename_for_log, log_accumulator) reward = get_image_reward(img, filename_for_log, log_accumulator) anime_aes_deepghs = get_anime_aesthetic_score_deepghs(img, filename_for_log, log_accumulator) maniqa = get_maniqa_score(img, filename_for_log, log_accumulator) clip_val = calculate_clip_score_value(img, prompt, filename_for_log, log_accumulator) sdxl_detect = get_sdxl_detection_score(img, filename_for_log, log_accumulator) anime_ai_chk_deepghs = get_anime_ai_check_score_deepghs(img, filename_for_log, log_accumulator) current_img_total_time_end = time.time() log_accumulator.append(f"INFO [{filename_for_log}]: Finished all scores (total for image: {current_img_total_time_end - current_img_total_time_start:.2f}s)") all_results.append({ "Filename": filename_for_log, "Prompt": prompt if prompt else "N/A", "Model Name": model_n, "Model Hash": model_h, "ImageReward": reward, "AnimeAesthetic_dg": anime_aes_deepghs, "MANIQA_TQ": maniqa, "CLIPScore": clip_val, "SDXL_Detector_AI_Prob": sdxl_detect, "AnimeAI_Check_dg_Prob": anime_ai_chk_deepghs, }) df_so_far = pd.DataFrame(all_results) progress( (i + 1.0) / len(files), desc=f"Img {i+1}/{len(files)}: Done {filename_for_log}") yield (df_so_far, gr.Image(visible=False), gr.Image(visible=False), gr.File(visible=False), gr.File(visible=False), f"Processed image {i+1}/{len(files)}: {filename_for_log}", "\n".join(log_accumulator)) except Exception as e: log_accumulator.append(f"CRITICAL ERROR processing {filename_for_log}: {e}") print(f"CRITICAL ERROR processing {filename_for_log}: {e}") all_results.append({ "Filename": filename_for_log, "Prompt": "Critical Error", "Model Name": "Error", "Model Hash": "Error", "ImageReward": "Error", "AnimeAesthetic_dg": "Error", "MANIQA_TQ": "Error", "CLIPScore": "Error", "SDXL_Detector_AI_Prob": "Error", "AnimeAI_Check_dg_Prob": "Error" }) df_so_far = pd.DataFrame(all_results) yield (df_so_far, gr.Image(visible=False), gr.Image(visible=False), gr.File(visible=False), gr.File(visible=False), f"Error on image {i+1}/{len(files)}: {filename_for_log}", "\n".join(log_accumulator)) log_accumulator.append("--- Generating final plots and download files ---") progress(1.0, desc="Generating final plots...") yield (pd.DataFrame(all_results), gr.Image(visible=False), gr.Image(visible=False), gr.File(visible=False), gr.File(visible=False), "Generating final plots...", "\n".join(log_accumulator)) df = pd.DataFrame(all_results) plot_model_avg_scores_buffer, plot_prompt_clip_scores_buffer = None, None csv_file_path_out, json_file_path_out = None, None if not df.empty: numeric_cols = ["ImageReward", "AnimeAesthetic_dg", "MANIQA_TQ", "CLIPScore"] for col in numeric_cols: df[col] = pd.to_numeric(df[col], errors='coerce') df_model_plot = df[(df["Model Name"] != "N/A") & (df["Model Name"].notna())] if not df_model_plot.empty and df_model_plot["Model Name"].nunique() > 0: try: model_avg_scores = df_model_plot.groupby("Model Name")[numeric_cols].mean().dropna(how='all') if not model_avg_scores.empty: fig1, ax1 = plt.subplots(figsize=(12, 7)); model_avg_scores.plot(kind="bar", ax=ax1) ax1.set_title("Average Scores per Model"); ax1.set_ylabel("Average Score") ax1.tick_params(axis='x', rotation=45, labelsize=8); plt.tight_layout() plot_model_avg_scores_buffer = io.BytesIO(); fig1.savefig(plot_model_avg_scores_buffer, format="png"); plot_model_avg_scores_buffer.seek(0); plt.close(fig1) log_accumulator.append("INFO: Model average scores plot generated.") except Exception as e: log_accumulator.append(f"ERROR: Failed to generate model average scores plot: {e}") df_prompt_plot = df[(df["Prompt"] != "N/A") & (df["Prompt"].notna()) & (df["CLIPScore"].notna())] if not df_prompt_plot.empty and df_prompt_plot["Prompt"].nunique() > 0 : try: df_prompt_plot["Short Prompt"] = df_prompt_plot["Prompt"].apply(lambda x: (str(x)[:30] + '...') if len(str(x)) > 33 else str(x)) prompt_clip_scores = df_prompt_plot.groupby("Short Prompt")["CLIPScore"].mean().sort_values(ascending=False) if not prompt_clip_scores.empty and len(prompt_clip_scores) >= 1 : fig2, ax2 = plt.subplots(figsize=(12, max(7, min(len(prompt_clip_scores)*0.5, 15)))) prompt_clip_scores.head(20).plot(kind="barh", ax=ax2) ax2.set_title("Average CLIPScore per Prompt (Top 20 unique prompts)"); ax2.set_xlabel("Average CLIPScore") plt.tight_layout(); plot_prompt_clip_scores_buffer = io.BytesIO(); fig2.savefig(plot_prompt_clip_scores_buffer, format="png"); plot_prompt_clip_scores_buffer.seek(0); plt.close(fig2) log_accumulator.append("INFO: Prompt CLIP scores plot generated.") except Exception as e: log_accumulator.append(f"ERROR: Failed to generate prompt CLIP scores plot: {e}") try: with tempfile.NamedTemporaryFile(mode="w+", delete=False, suffix=".csv", encoding='utf-8') as tmp_csv: df.to_csv(tmp_csv, index=False); csv_file_path_out = tmp_csv.name with tempfile.NamedTemporaryFile(mode="w+", delete=False, suffix=".json", encoding='utf-8') as tmp_json: df.to_json(tmp_json, orient='records', indent=4); json_file_path_out = tmp_json.name log_accumulator.append("INFO: CSV and JSON data prepared for download.") except Exception as e: log_accumulator.append(f"ERROR preparing download files: {e}") final_status = f"Finished processing {len(all_results)} images." log_accumulator.append(final_status) # Преобразуем BytesIO в PIL.Image перед передачей в gr.Image pil_plot_model_avg = Image.open(plot_model_avg_scores_buffer) if plot_model_avg_scores_buffer and plot_model_avg_scores_buffer.getbuffer().nbytes > 0 else None pil_plot_prompt_clip = Image.open(plot_prompt_clip_scores_buffer) if plot_prompt_clip_scores_buffer and plot_prompt_clip_scores_buffer.getbuffer().nbytes > 0 else None if pil_plot_model_avg or pil_plot_prompt_clip: log_accumulator.append("INFO: Plots converted to PIL Images for display.") else: log_accumulator.append("INFO: No plots were generated or plots are empty.") yield ( df, gr.Image(value=pil_plot_model_avg, visible=pil_plot_model_avg is not None), gr.Image(value=pil_plot_prompt_clip, visible=pil_plot_prompt_clip is not None), gr.File(value=csv_file_path_out, visible=csv_file_path_out is not None), gr.File(value=json_file_path_out, visible=json_file_path_out is not None), final_status, "\n".join(log_accumulator) ) with gr.Blocks(css="footer {display: none !important}") as demo: gr.Markdown("# AI Image Model Evaluation Tool") gr.Markdown("Upload PNG images (ideally with Stable Diffusion metadata) to evaluate them...") with gr.Row(): image_uploader = gr.Files(label="Upload Images (PNG)", file_count="multiple", file_types=["image"]) process_button = gr.Button("Evaluate Images", variant="primary") status_textbox = gr.Textbox(label="Overall Status", interactive=False) log_output_textbox = gr.Textbox(label="Detailed Logs", lines=15, interactive=False, autoscroll=True) gr.Markdown("## Evaluation Results Table") results_table = gr.DataFrame(headers=[ "Filename", "Prompt", "Model Name", "Model Hash", "ImageReward", "AnimeAesthetic_dg", "MANIQA_TQ", "CLIPScore", "SDXL_Detector_AI_Prob", "AnimeAI_Check_dg_Prob" ], wrap=True) with gr.Row(): download_csv_button = gr.File(label="Download CSV Results", interactive=False) download_json_button = gr.File(label="Download JSON Results", interactive=False) gr.Markdown("## Visualizations") with gr.Row(): plot_output_model_avg = gr.Image(label="Average Scores per Model", type="pil", interactive=False) plot_output_prompt_clip = gr.Image(label="Average CLIPScore per Prompt", type="pil", interactive=False) process_button.click( fn=process_images_generator, inputs=[image_uploader], outputs=[results_table, plot_output_model_avg, plot_output_prompt_clip, download_csv_button, download_json_button, status_textbox, log_output_textbox] ) gr.Markdown("""**Metric Explanations:** ... (без изменений)""") if __name__ == "__main__": print("--- Initializing models, please wait... ---") initial_dummy_logs = [] if onnxruntime: get_onnx_session_and_meta(ANIME_AESTHETIC_REPO, ANIME_AESTHETIC_SUBFOLDER, initial_dummy_logs) get_onnx_session_and_meta(ANIME_AI_CHECK_REPO, ANIME_AI_CHECK_SUBFOLDER, initial_dummy_logs) if initial_dummy_logs: print("--- Initial ONNX loading attempts log: ---") for log_line in initial_dummy_logs: print(log_line) print("-----------------------------------------") print("--- Model initialization attempt complete. Launching Gradio. ---") demo.queue().launch(debug=True)