v9: Gemini 3 Pro Image primary + 3.1 Flash + 2.5 Flash fallback, improved prompts

app.py

@@ -7,7 +7,9 @@ GEMINI_API_KEY = os.environ.get('GEMINI_API_KEY', '')
 HF_TOKEN       = os.environ.get('HF_TOKEN', '')
 
 GEMINI_TEXT_URL  = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent'
-GEMINI_IMAGE_URL = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent'
+GEMINI_IMAGE_URL  = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent'
+GEMINI_IMAGE_URL2 = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-image-preview:generateContent'
+GEMINI_IMAGE_URL3 = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent'
 
 # ── FONTS ─────────────────────────────────────────────────────────────────────
 _B = ['/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf',
@@ -37,9 +39,8 @@ def gemini_text(prompt, api_key):
     return data['candidates'][0]['content']['parts'][0]['text']
 
 # ── GEMINI IMAGEN (page generation) ──────────────────────────────────────────
-def gemini_imagen(prompt, api_key, width=1024, height=1408):
-    """Gemini 2.0 Flash Preview Image Generation - free tier compatible."""
-    url = GEMINI_IMAGE_URL + '?key=' + api_key
+def gemini_imagen_call(url, prompt, api_key, width, height):
+    """Single Gemini image model call with retry on 429."""
     body = {
         'contents': [{'parts': [{'text': prompt}]}],
         'generationConfig': {
@@ -47,9 +48,10 @@ def gemini_imagen(prompt, api_key, width=1024, height=1408):
             'temperature': 1.0
         }
     }
-    for attempt in range(3):
+    full_url = url + '?key=' + api_key
+    for attempt in range(2):
         try:
-            r = requests.post(url, json=body, timeout=120)
+            r = requests.post(full_url, json=body, timeout=120)
             if r.status_code == 200:
                 data = r.json()
                 candidates = data.get('candidates', [])
@@ -60,17 +62,41 @@ def gemini_imagen(prompt, api_key, width=1024, height=1408):
                             img_bytes = base64.b64decode(part['inlineData']['data'])
                             img = Image.open(BytesIO(img_bytes)).convert('RGB')
                             return img.resize((width, height), Image.LANCZOS)
+                print('Gemini image: no image in response from ' + url.split('/models/')[1].split(':')[0])
+                return None
             elif r.status_code == 429:
-                print('Gemini image 429, waiting ' + str(15*(attempt+1)) + 's...')
-                time.sleep(15 * (attempt + 1))
+                print('Gemini image 429 (' + url.split('/models/')[1].split(':')[0] + '), waiting ' + str(20*(attempt+1)) + 's...')
+                time.sleep(20 * (attempt + 1))
             else:
-                print('Gemini image error ' + str(r.status_code) + ': ' + r.text[:300])
+                print('Gemini image error ' + str(r.status_code) + ' (' + url.split('/models/')[1].split(':')[0] + '): ' + r.text[:200])
                 return None
         except Exception as e:
             print('Gemini image exception: ' + str(e)[:100])
             return None
     return None
 
+def gemini_imagen(prompt, api_key, width=1024, height=1408):
+    """Try Gemini image models from best to fastest: 3-pro → 3.1-flash → 2.5-flash."""
+    # Model 1: Gemini 3 Pro Image (best quality)
+    print('  Trying Gemini 3 Pro Image...')
+    img = gemini_imagen_call(GEMINI_IMAGE_URL, prompt, api_key, width, height)
+    if img:
+        print('  Gemini 3 Pro Image SUCCESS')
+        return img
+    # Model 2: Gemini 3.1 Flash Image (fast + good quality)
+    print('  Trying Gemini 3.1 Flash Image...')
+    img = gemini_imagen_call(GEMINI_IMAGE_URL2, prompt, api_key, width, height)
+    if img:
+        print('  Gemini 3.1 Flash Image SUCCESS')
+        return img
+    # Model 3: Gemini 2.5 Flash Image (fastest)
+    print('  Trying Gemini 2.5 Flash Image...')
+    img = gemini_imagen_call(GEMINI_IMAGE_URL3, prompt, api_key, width, height)
+    if img:
+        print('  Gemini 2.5 Flash Image SUCCESS')
+        return img
+    return None
+
 # ── POLLINATIONS FALLBACK ─────────────────────────────────────────────────────
 def pollinations(prompt, seed, retries=3):
     enc = urllib.parse.quote(prompt[:600])
@@ -182,33 +208,66 @@ def fallback_script(topic, num_pages, style, language):
 
 # ── PAGE IMAGE PROMPT BUILDER ─────────────────────────────────────────────────
 def build_page_prompt(page_data, script, style):
-    chars = script.get('characters', [])
-    char_ref = '. '.join(c['name'] + ': ' + c['description'] for c in chars[:2])
-    panels = page_data.get('panels', [])[:4]
-    base = page_data.get('page_image_prompt', '')
-
-    if not base or len(base) < 50:
-        # Build from panels
-        parts = []
-        for i, p in enumerate(panels):
-            dlg_text = ' | '.join('"' + d['text'] + '" - ' + d['character']
-                                  for d in p.get('dialogue', [])[:1])
-            parts.append('Panel ' + str(i+1) + ': ' + str(p.get('scene','')) +
-                         ', ' + str(p.get('action','')) +
-                         ('. Speech bubble: ' + dlg_text if dlg_text else ''))
-        base = '. '.join(parts)
+    """Build a rich, detailed prompt for Gemini 3 Pro image generation."""
+    title     = script.get('title', 'Comic')
+    genre     = script.get('genre', 'adventure')
+    chars     = script.get('characters', [])
+    panels    = page_data.get('panels', [])[:4]
+    page_num  = page_data.get('page_number', 1)
+    total_p   = script.get('total_pages', 1)
+
+    # Character descriptions for consistency
+    char_descs = []
+    for c in chars[:3]:
+        char_descs.append(c['name'] + ': ' + c.get('description','') + (', wears ' + c.get('costume','') if c.get('costume') else ''))
+    char_ref = '. '.join(char_descs) if char_descs else ''
+
+    # Build per-panel descriptions with full dialogue
+    panel_parts = []
+    num_panels = len(panels)
+    for i, p in enumerate(panels):
+        scene   = str(p.get('scene', ''))
+        action  = str(p.get('action', ''))
+        caption = str(p.get('caption', ''))
+        dialogues = p.get('dialogue', [])
+        dlg_list = []
+        for d in dialogues:
+            dlg_list.append(d['character'] + ' says: "' + d['text'] + '"')
+        dlg_str = '; '.join(dlg_list) if dlg_list else 'no dialogue'
+        cap_str = (' Narrator caption box: "' + caption + '"') if caption else ''
+        panel_parts.append(
+            'PANEL ' + str(i+1) + ': Scene=' + scene +
+            '. Action=' + action + '. ' + dlg_str + cap_str
+        )
+
+    # Choose layout based on panel count
+    if num_panels == 1:
+        layout_desc = 'ONE large full-page panel'
+    elif num_panels == 2:
+        layout_desc = 'TWO panels: large top panel, large bottom panel, separated by thick black gutter'
+    elif num_panels == 3:
+        layout_desc = 'THREE panels: wide panoramic top panel, two equal panels side-by-side at bottom, separated by thick black gutters'
+    else:
+        layout_desc = 'FOUR panels: large wide panel at top, two medium square panels in middle row, one wide panel at bottom, all separated by thick black gutters (8px)'
+
+    panels_text = ' | '.join(panel_parts)
 
     prompt = (
-        style + ' comic book FULL PAGE, professional comic art, high detail. '
-        'LAYOUT: 4 panels arranged with clear black borders and white gutters between them. '
-        'Panel arrangement: large panel top, two medium panels middle, one wide panel bottom. '
-        + str(base) + '. '
-        'IMPORTANT: Each panel must have white oval speech bubbles with BLACK TEXT showing character dialogue. '
-        'Bubbles should be small (max 20% of panel area), positioned near speaking character. '
-        'Characters consistent: ' + char_ref + '. '
-        'Comic book style: bold outlines, flat vivid colors, dramatic shadows, NO watermarks.'
+        'CREATE A PROFESSIONAL ' + style.upper() + ' COMIC BOOK PAGE. '
+        'Comic: "' + title + '" (' + genre + '). Page ' + str(page_num) + ' of ' + str(total_p) + '. '
+        'LAYOUT: ' + layout_desc + '. '
+        'PANELS: ' + panels_text + '. '
+        'SPEECH BUBBLES: White rounded oval bubbles with clear BLACK TEXT for every dialogue line. '
+        'Bubbles sized to fit text (max 25% of panel area), tail pointing to speaker. '
+        'Text must be legible. Do NOT overlap speech bubbles with main action. '
+        'STYLE REQUIREMENTS: '
+        + style + ' art style. Bold black panel borders (8-10px). White gutter between panels. '
+        'Dynamic compositions. Dramatic lighting and shadows. Vivid expressive colors. '
+        'Professional comic book quality. NO watermarks, NO logos, NO page numbers in image. '
+        + (('CHARACTERS (keep consistent): ' + char_ref + '. ') if char_ref else '') +
+        'Make every panel visually compelling and clearly readable as a comic book.'
     )
-    return prompt[:1000]
+    return prompt[:1800]
 
 # ── GENERATE ONE PAGE ─────────────────────────────────────────────────────────
 def generate_page(page_data, script, style, title, page_num, total, api_key):
@@ -217,7 +276,7 @@ def generate_page(page_data, script, style, title, page_num, total, api_key):
 
     print('Page ' + str(page_num) + '/' + str(total) + ' generating...')
 
-    # Try Gemini 2.0 Flash Image first
+    # Try Gemini image models (3-pro → 3.1-flash → 2.5-flash → Pollinations)
     art = None
     if api_key:
         art = gemini_imagen(prompt, api_key)
@@ -275,7 +334,7 @@ def generate_comic_ui(topic, num_pages, style, language, gemini_key):
     script, pages = generate_comic(topic, num_pages, style, language, gemini_key)
     info = json.dumps({'title':script.get('title',''),'summary':script.get('story_summary',''),
                        'characters':script.get('characters',[]),'total_pages':len(pages),
-                       'engine':'Gemini 2.0 Flash Image → Pollinations FLUX fallback'},
+                       'engine':'Gemini 3 Pro Image → 3.1 Flash → 2.5 Flash → Pollinations fallback'},
                       ensure_ascii=False, indent=2)
     return script.get('title',topic), script.get('story_summary',''), pages, info
 
@@ -284,7 +343,7 @@ STYLES=['manga','superhero comic','watercolor comic','noir comic','cartoon',
 
 with gr.Blocks(title='Comic AI Generator') as demo:
     gr.Markdown('# Comic AI Generator')
-    gr.Markdown('Powered by **Gemini 2.5 Flash Image** — each page is one AI-generated full comic page (1240×1754 px) with panels, borders and speech bubbles drawn by AI.')
+    gr.Markdown('Powered by **Gemini 3 Pro Image** — each page is one AI-generated full comic page (1240×1754 px) with panels, borders and speech bubbles drawn by AI.')
     with gr.Tabs():
         with gr.Tab('Generate Comic'):
             with gr.Row():