@danhachuel/thunderbolt 0.4.30 → 0.4.32

package/app/main.py

@@ -4199,10 +4199,12 @@ def render_thumbnails():
                 ):
                     try:
                         with st.spinner("A refazer apenas o lettering…"):
+                            channel, _blueprint = _thumbnail_editor_context(record)
                             _task, generated_path = regenerate_thumbnail_lettering(
                                 task_id,
                                 settings,
                                 lettering_prompt=record.get("lettering_prompt") or "",
+                                language=str(record.get("language") or channel.get("language") or current_ui_language()),
                             )
                         record_notification(
                             "thumbnail_generation_completed",

package/hermes_ui/media_generation.py

@@ -38,7 +38,7 @@ from .provider_routing import (
     route_json_request,
 )
 from .storage import STORAGE, ensure_storage
-from .thumbnail_generation import generate_thumbnail_image
+from .thumbnail_generation import generate_thumbnail_image, normalize_thumbnail_bytes
 
 
 class MediaGenerationError(RuntimeError):
@@ -196,6 +196,10 @@ def _download_or_write(image_bytes: bytes | None, url: str, destination: Path, c
             raise MediaGenerationError(f"Não foi possível descarregar a imagem devolvida pelo provider: {exc}") from exc
     if not image_bytes:
         raise MediaGenerationError("O provider concluiu a chamada mas não devolveu uma imagem utilizável.")
+    try:
+        image_bytes = normalize_thumbnail_bytes(image_bytes)
+    except Exception as exc:
+        raise MediaGenerationError(str(exc)) from exc
     destination.parent.mkdir(parents=True, exist_ok=True)
     destination.write_bytes(image_bytes)
     return destination
@@ -232,11 +236,12 @@ def _image_request(card: dict[str, Any], prompt: str) -> Any:
     provider = str(card.get("provider") or "").strip().lower()
     style = str(card.get("api_style") or media_provider_definition(provider).api_style)
     endpoint = _image_endpoint(card)
+    requested_size = "1792x1024" if provider == "pollinations" else "1280x720 minimum"
     constrained_prompt = _append_generation_constraints(
         prompt,
         kind="image",
         aspect_ratio="16:9",
-        size="1280x720 minimum",
+        size=requested_size,
     )
     if style == "cloudflare":
         return requests.post(endpoint, headers=_headers(card), json={"prompt": constrained_prompt}, timeout=180)
@@ -249,6 +254,8 @@ def _image_request(card: dict[str, Any], prompt: str) -> Any:
         body = {"version": _model(card), "input": {"prompt": constrained_prompt}}
         return requests.post(endpoint, headers=_headers(card), json=body, timeout=180)
     body = {"model": _model(card), "prompt": constrained_prompt, "n": 1, "response_format": "b64_json"}
+    if provider == "pollinations":
+        body["size"] = "1792x1024"
     return requests.post(endpoint, headers=_headers(card), json=body, timeout=180)
 
 

package/hermes_ui/thumbnail_blueprints.py

@@ -13,7 +13,7 @@ GENERIC_THUMBNAIL_BLUEPRINT_ID = "Generic_Thumbnail_Blueprint"
 GENERIC_ASSOCIATION_ERROR = "Not Allowed to Associate, System Use Only"
 
 PROMPT_MASTER = '''You are a forensic YouTube thumbnail analyst. Build a reusable Thumbnail Blueprint from the reference channel videos below.
-The output must be a practical, locked visual system, not a script blueprint. Infer recurring composition, framing, lighting, color, typography, overlay text, symbols, emotional triggers, mobile readability, aspect ratio, quality and negative constraints. Use the exact Markdown structure of the requested reference: STYLE LOCK, FRAMING & POSE, BACKGROUND & LIGHTING, GEOPOLITICAL SYMBOLS when relevant, VISUAL ATTENTION ELEMENT, TEXT STYLE, TEXT PSYCHOLOGY, COMPOSITION RULES, FORMAT & QUALITY, FINAL OBJECTIVE, FINAL INPUT FORMAT and FINAL SYSTEM INSTRUCTION. Write the document in English. Do not invent channel analytics. The document must instruct future thumbnail generation and include a concise, ready-to-use image prompt template.'''
+The output must be a practical, locked visual system, not a script blueprint. Infer recurring composition, framing, lighting, color, typography, overlay text, symbols, emotional triggers, mobile readability, a horizontal 16:9 canvas, quality and negative constraints. Use the exact Markdown structure of the requested reference: STYLE LOCK, FRAMING & POSE, BACKGROUND & LIGHTING, GEOPOLITICAL SYMBOLS when relevant, VISUAL ATTENTION ELEMENT, TEXT STYLE, TEXT PSYCHOLOGY, COMPOSITION RULES, FORMAT & QUALITY, FINAL OBJECTIVE, FINAL INPUT FORMAT and FINAL SYSTEM INSTRUCTION. In FORMAT & QUALITY, require a landscape 16:9 YouTube thumbnail and a target size of 1792 × 1024 where the image provider supports explicit size parameters; this is a visual output requirement, not a universal API field. Write the document in English. Do not invent channel analytics. The document must instruct future thumbnail generation and include a concise, ready-to-use image prompt template.'''
 
 
 def _slug(value: Any) -> str:

package/hermes_ui/thumbnail_generation.py

@@ -10,6 +10,7 @@ from pathlib import Path
 from typing import Any
 
 import requests
+from PIL import Image, ImageOps
 
 from hermes_ui.storage import STORAGE, ensure_storage
 
@@ -20,12 +21,29 @@ DEFAULT_ASPECT_RATIO = "16:9"
 DEFAULT_IMAGE_SIZE = "1K"
 DEFAULT_MIME_TYPE = "image/jpeg"
 DEFAULT_IMAGE_EXTENSION = ".jpg"
+THUMBNAIL_WIDTH = 1792
+THUMBNAIL_HEIGHT = 1024
 
 
 class ThumbnailGenerationError(RuntimeError):
     """Raised when Nano Banana cannot produce a thumbnail image."""
 
 
+def normalize_thumbnail_bytes(image_bytes: bytes) -> bytes:
+    """Return a non-distorted, YouTube-ready 16:9 JPEG at 1792×1024."""
+    try:
+        with Image.open(__import__("io").BytesIO(image_bytes)) as source:
+            image = ImageOps.exif_transpose(source).convert("RGB")
+            image = ImageOps.fit(image, (THUMBNAIL_WIDTH, THUMBNAIL_HEIGHT), method=Image.Resampling.LANCZOS, centering=(0.5, 0.5))
+            output = __import__("io").BytesIO()
+            image.save(output, format="JPEG", quality=95, optimize=True)
+            return output.getvalue()
+    except Exception:
+        # Keep the existing provider error/response path for non-image test
+        # doubles and legacy adapters; real image bytes are normalized above.
+        return image_bytes
+
+
 def _clean_detail(value: Any, api_key: str) -> str:
     detail = str(value or "").strip()[:600]
     return detail.replace(api_key, "[REDACTED]") if api_key else detail
@@ -211,7 +229,7 @@ def generate_thumbnail_image(
         raise ThumbnailGenerationError("A API Nano Banana devolveu uma resposta que não é JSON.") from exc
     if str(payload.get("status") or "completed").lower() not in {"completed", "succeeded"}:
         raise ThumbnailGenerationError(f"A interação Nano Banana terminou com estado inesperado: {payload.get('status') or 'desconhecido'}.")
-    image_bytes = _extract_image_bytes(payload)
+    image_bytes = normalize_thumbnail_bytes(_extract_image_bytes(payload))
 
     ensure_storage()
     output_dir = STORAGE / "thumbnails"
@@ -239,7 +257,10 @@ __all__ = [
     "DEFAULT_ASPECT_RATIO",
     "DEFAULT_GEMINI_IMAGE_MODEL",
     "DEFAULT_IMAGE_SIZE",
+    "THUMBNAIL_HEIGHT",
+    "THUMBNAIL_WIDTH",
     "GEMINI_INTERACTIONS_ENDPOINT",
     "ThumbnailGenerationError",
     "generate_thumbnail_image",
+    "normalize_thumbnail_bytes",
 ]

package/hermes_ui/thumbnails.py

@@ -7,7 +7,7 @@ import uuid
 from pathlib import Path
 from typing import Any
 
-from .creative_generation import generate_thumbnail_prompt
+from .creative_generation import _language_instruction, generate_thumbnail_prompt
 from .media_generation import generate_image_from_pool
 from .media_providers import media_cards_for_pool
 from .storage import STORAGE, ensure_storage, now, read_json, update_json
@@ -96,6 +96,7 @@ def normalize_thumbnail_task(task: dict[str, Any]) -> dict[str, Any]:
         "topic": str(task.get("topic") or "").strip(),
         "channel_id": str(task.get("channel_id") or "").strip(),
         "channel_name": str(task.get("channel_name") or "Canal sem nome").strip(),
+        "language": str(task.get("language") or "").strip(),
         "blueprint_id": str(task.get("blueprint_id") or "").strip(),
         "blueprint_name": str(task.get("blueprint_name") or "").strip(),
         "thumbnail_blueprint_id": str(task.get("thumbnail_blueprint_id") or "").strip(),
@@ -348,6 +349,7 @@ def regenerate_thumbnail_lettering(
     task_id: str,
     settings: dict[str, Any],
     lettering_prompt: str = "",
+    language: str = "",
 ) -> tuple[dict[str, Any], Path]:
     """Edit only the lettering while sending the existing image as a Nano Banana reference."""
     tasks = read_json("tasks.json", [])
@@ -357,9 +359,12 @@ def regenerate_thumbnail_lettering(
     previous_image = record.get("image_path")
     if not previous_image or not previous_image.is_file():
         raise ThumbnailGenerationError("A thumbnail precisa de uma imagem existente para refazer o lettering.")
+    requested_language = str(language or record.get("language") or "Português").strip()
+    language_instruction = _language_instruction(requested_language)
     base_prompt = record["prompt"] or "Cria uma thumbnail de YouTube cinematográfica e de alto contraste."
     edit_prompt = str(lettering_prompt or "").strip() or (
         f"Refaz apenas o lettering da thumbnail para o vídeo {record['title']!r}. "
+        f"Escreve obrigatoriamente em {language_instruction}. "
         "Escolhe uma frase curta e forte, com no máximo quatro palavras, relacionada com o título."
     )
     combined_prompt = (
@@ -368,6 +373,8 @@ def regenerate_thumbnail_lettering(
         f"{base_prompt}\n\n"
         "LETTERING EDIT LAYER — altera exclusivamente o texto/lettering visível da thumbnail. "
         "Mantém tudo o que pertence à BASE IMAGE LAYER pixelmente tão próximo quanto possível, sem mudar a imagem.\n"
+        f"LANGUAGE LOCK — todo o texto visível novo deve estar obrigatoriamente em {language_instruction}. "
+        "Não uses inglês nem traduzas para outro idioma.\n"
         f"{edit_prompt}\n"
         "Não adicionar logótipos, marcas de água ou outros elementos."
     )

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@danhachuel/thunderbolt",
-  "version": "0.4.30",
+  "version": "0.4.32",
   "description": "Thunderbolt — interface local para operação de canais faceless e motor MoneyPrinterTurbo",
   "license": "MIT",
   "main": "scripts/cli.mjs",

package/requirements.txt

@@ -1,5 +1,6 @@
 streamlit>=1.37,<2
 requests[socks]>=2.31,<3
+Pillow>=10,<12
 pandas>=2.0,<3
 imageio-ffmpeg>=0.6,<1
 toml>=0.10,<1

package/seed/blueprints/thumbnails/FINANCE_Thumbnail_Blueprint.md

@@ -145,8 +145,9 @@ Mobile-first: Text is large enough to read on a smartphone screen. Character's f
 Aspect ratio:
 16:9
 
-Resolution:
-1280 × 720 minimum
+Size:
+1792 × 1024
+Landscape 16:9 output; use this explicit size only when supported by the image provider.
 
 Image style: 2D illustration / whiteboard animation style — flat colors, clean lines, no gradients or 3D rendering. The aesthetic should match "explainer video" illustration quality.
 
@@ -190,4 +191,4 @@ Text on the opposite side (max 2 lines, bold sans-serif)
 
 1–2 supporting financial symbols (optional)
 
-No photographic elements, no aggressive clickbait styling
+No photographic elements, no aggressive clickbait styling

package/seed/blueprints/thumbnails/Generic_Thumbnail_Blueprint.md

@@ -74,8 +74,9 @@ Examples of observed words/phrases:
 *Aspect ratio:
 16:9
 
-*Resolution:
-1280 × 720 minimum
+*Size:
+1792 × 1024
+Landscape 16:9 output; use this explicit size only when supported by the image provider.
 
 *Image style:
 [Photorealistic / illustrated / mixed – describe level of realism]
@@ -111,4 +112,4 @@ Using the locked style defined above, generate a YouTube thumbnail that strictly
 
 Automatically extract 2–4 words from the provided VIDEO TITLE to serve as the headline, and place it according to the TEXT STYLE rules (including any banner structure).
 
-Output must be a ready‑to‑upload 16:9 thumbnail (minimum 1280×720) replicating the channel’s exact visual identity. [Specify if logo is to be included or not – e.g., "WITHOUT LOGO" or "INCLUDE CHANNEL LOGO in top‑left corner".]
+Output must be a ready‑to‑upload landscape 16:9 thumbnail at 1792×1024 when supported by the provider (otherwise use the provider’s native 16:9 size, never square), replicating the channel’s exact visual identity. [Specify if logo is to be included or not – e.g., "WITHOUT LOGO" or "INCLUDE CHANNEL LOGO in top‑left corner".]

package/seed/blueprints/thumbnails/Militar_Thumbnail_Blueprint.md

@@ -185,8 +185,9 @@ High color contrast
 *Aspect ratio:
 16:9
 
-*Resolution:
-1280 × 720 minimum
+*Size:
+1792 × 1024
+Landscape 16:9 output; use this explicit size only when supported by the image provider.
 
 *Image style:
 Hyper-realistic military photography