Try to improve prompt adherence for low variety image models

vibe_bot/main.py

@@ -146,13 +146,17 @@ IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE = (
     "The final image will use a {layout} canvas, so compose the scene to fit "
     "that orientation.\n"
     "Your prompt must cover all of the following as one flowing, descriptive "
-    "passage:\n"
+    "passage. Begin with the main subject, described completely in the very "
+    "first sentence:\n"
     "- Subject(s): every subject with concrete specifics (species or "
     "character, age, build, clothing, colors, materials, accessories), its "
     "exact pose, expression, gaze direction, and its precise position in the "
     'frame (for example "centered in the foreground" or "small in the '
     'upper-left background"). State the relative scale of subjects to each '
-    "other and to the frame.\n"
+    "other and to the frame. If the subject is a fusion, hybrid, or anything "
+    "unusual, the first sentence must state in full what is joined to what "
+    "and exactly how it looks, and that description must be repeated near "
+    "the end of the passage.\n"
     "- Composition and framing: the camera angle (eye-level, low, high, "
     "bird's-eye), the shot type (extreme close-up, portrait, full body, wide "
     "establishing shot), the focal point, the arrangement of elements across "
@@ -177,8 +181,23 @@ IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE = (
     '- Be concrete and specific. Never use vague words like "nice", '
     '"cool", "epic", or "various" - name exact colors, objects, '
     "positions, and quantities.\n"
-    "- If the request is vague or incomplete, fill in the missing details "
-    "with coherent choices that fit the request.\n"
+    "- Be literal. Interpret the user's request exactly as written. Never "
+    'rationalize, normalize, or "improve" it: surreal, absurd, or '
+    'anthropomorphic requests are intentional, not mistakes. A "fountain '
+    'pen wearing pants" is an anthropomorphized fountain pen character '
+    "wearing pants, not a pen lying next to a pair of pants.\n"
+    "- Preserve everything the user specified. Fill in details the user did "
+    "not specify with coherent choices that fit the request, but never "
+    "alter, drop, or reinterpret what the user did specify.\n"
+    "- Decompose concepts. The image model lacks world knowledge, so never "
+    "rely on a name alone for anything it might misrender (mythical "
+    "creatures, fictional characters, cultural items, animal breeds, "
+    "instruments, vehicles). Spell out the visual anatomy: silhouette, body "
+    "parts, materials, and distinguishing features, with explicit "
+    "disambiguation. A centaur is a single creature with a human torso, "
+    "arms, and head seamlessly fused to a horse's front half, the horse's "
+    "four legs extending from the human's waist - one fused body, not a "
+    "person riding a horse.\n"
     "- If told to generate an image of yourself, generate a picture of a "
     "canada goose. If told to generate a picture of 'me', 'myself', or some "
     "other self reference, generate a picture of a canada goose.\n"
@@ -228,6 +247,54 @@ def select_image_layout(user_message: str) -> str:
     return parse_image_layout(response)
 
 
+IMAGE_PROMPT_VERIFY_SYSTEM_PROMPT = (
+    "You are the final quality check for an image generation prompt. The "
+    "image model that will use it has no world knowledge: it renders only "
+    "what is described literally and silently drops anything it does not "
+    "understand - a prompt that merely names a centaur without describing "
+    "the fused human-animal body will produce a plain horse. "
+    "Given the user's original request and the drafted prompt, judge "
+    "strictly: would the drafted prompt, taken completely literally, "
+    "produce exactly what the user asked for, including every unusual, "
+    "mythical, surreal, or anthropomorphic element? "
+    "Respond with ONLY the single word PASS if it would. Otherwise respond "
+    "with ONLY a corrected version of the prompt that would produce exactly "
+    "what the user asked for: one flowing descriptive passage, the "
+    "subject's full anatomy and every unusual element described explicitly "
+    "in the first sentence and repeated near the end, no other text."
+)
+
+
+def verify_image_prompt(user_message: str, image_prompt: str) -> str:
+    """Check the drafted prompt literally produces the user's request.
+
+    Args:
+        user_message: The user's original image request.
+        image_prompt: The drafted image generation prompt.
+
+    Returns:
+        The original prompt when the check passes or the LLM returns an
+        empty response, otherwise the LLM's corrected prompt.
+
+    """
+    check_prompt = f"User request: {user_message}\n\nDrafted prompt: {image_prompt}"
+    response = llama_wrapper.chat_completion_instruct(
+        system_prompt=IMAGE_PROMPT_VERIFY_SYSTEM_PROMPT,
+        user_prompt=check_prompt,
+        openai_url=CHAT_ENDPOINT,
+        openai_api_key=CHAT_ENDPOINT_KEY,
+        model=CHAT_MODEL,
+        max_tokens=MAX_COMPLETION_TOKENS,
+    )
+    if not response:
+        return image_prompt
+    # A passing check is the single word PASS; a correction is a full
+    # rewritten passage, which is always much longer.
+    if len(response) <= 50 and re.search(r"\bpass\b", response, re.IGNORECASE):
+        return image_prompt
+    return response
+
+
 @bot.event
 async def on_ready() -> None:
     """Log when the bot is ready and logged in."""
@@ -862,6 +929,11 @@ async def doodlebob(ctx: CommandsContext[Bot], *, message: str) -> None:
         logger.warning("No image prompt supplied. Check for errors.")
         return
 
+    # Verify the prompt literally produces the user's request; the check
+    # may return a corrected prompt.
+    image_prompt = verify_image_prompt(message, image_prompt)
+    logger.info("Doodlebob final image prompt: %s", image_prompt)
+
     # Alert the user we're generating the image
     db = get_database()
     estimated_seconds = db.get_image_generation_time_estimate()