improve image gen prompt

2026-08-17 12:54:31 -04:00
parent b740d00b5d
commit 7f0c5fbd80
2 changed files with 99 additions and 10 deletions
@@ -135,6 +135,58 @@ IMAGE_LAYOUT_SYSTEM_PROMPT = (
    "- Do NOT include any other text, punctuation, explanation, or reasoning.\n"
 )

+IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE = (
+    "You are an expert art director and image-generation prompt engineer. "
+    "Convert the user's message into one single, extremely detailed image "
+    "generation prompt that will be passed directly to a text-to-image model. "
+    "The image model is weak: it guesses at compositions, fumbles rendered "
+    "text, and invents details on its own. Your prompt must therefore leave "
+    "nothing to interpretation - explicitly describe every visible aspect of "
+    "the image so it can be created with extreme precision and detail.\n"
+    "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"
+    "- 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"
+    "- 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 "
+    "the {layout} canvas, and the depth of field.\n"
+    "- Text: if the image must contain readable text (titles, signs, "
+    "posters, labels, banners, watermarks, logos, captions), quote the EXACT "
+    "text verbatim in double quotes with precise capitalization and "
+    "punctuation, and specify its font style, color, size, and exact "
+    "placement. If the image should contain no text, state that explicitly "
+    '("no text anywhere in the image").\n'
+    "- Setting and background: the complete environment with concrete "
+    "details - location, time of day, weather, and every notable background "
+    "and foreground element with its position.\n"
+    "- Style and rendering: the art style or medium (for example "
+    "photorealistic 35mm photograph, oil painting, watercolor, cel-shaded "
+    "anime, pixel art, vector illustration), the color palette with specific "
+    "colors, the lighting (source, direction, quality, mood), the overall "
+    "atmosphere, and the level of detail.\n"
+    '- Finish with concise quality terms such as "highly detailed, sharp '
+    'focus".\n'
+    "Rules:\n"
+    '- 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"
+    "- 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"
+    "- Respond with ONLY the image generation prompt itself. Do not affirm "
+    "the user, do not answer the user's questions, and do not add headings, "
+    "labels, numbered lists, or any other text."
+)
+

 def parse_image_layout(response: str) -> str:
    """Parse an LLM response into a valid image layout.
@@ -793,16 +845,7 @@ async def doodlebob(ctx: CommandsContext[Bot], *, message: str) -> None:
    logger.info("Doodlebob selected layout %r for %s", layout, ctx.author.name)
    await ctx.send(f"**Doodlebob selected {layout}**")

-    system_prompt = (
-        "Given the following message, convert it to a detailed image generation "
-        "prompt that will be passed directly into an image generation model. "
-        f"The final image will use a {layout} canvas, so compose the scene to "
-        f"fit that orientation. "
-        "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. Only respond with a valid image "
-        "generation prompt, do not affirm the user or respond to the user's questions."
-    )
+    system_prompt = IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE.format(layout=layout)

    # Wait for the generated image prompt
    image_prompt = llama_wrapper.chat_completion_instruct(
@@ -1018,6 +1018,52 @@ def test_parse_image_layout_defaults_to_square(response: str) -> None:
    assert main_module.parse_image_layout(response) == "square"


+def test_image_prompt_system_prompt_covers_key_details() -> None:
+    """The prompt-rewrite system prompt forces explicit detail on all aspects."""
+    import vibe_bot.main as main_module
+
+    prompt = main_module.IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE.format(layout="square")
+    lowered = prompt.lower()
+    assert "square" in prompt
+    assert "exact text" in lowered
+    assert "composition" in lowered
+    assert "style" in lowered
+    assert "canada goose" in lowered
+    assert "only the image generation prompt" in lowered
+
+
+def test_doodlebob_prompt_rewrite_uses_detailed_system_prompt(
+    mock_ctx: MagicMock,
+    mock_llama_wrapper: MagicMock,
+    mock_base64: MagicMock,
+    mock_database: MagicMock,
+) -> None:
+    """The prompt rewrite call uses the detailed system prompt with the layout."""
+    import asyncio
+
+    import vibe_bot.main as main_module
+
+    mock_llama_wrapper.chat_completion_instruct.side_effect = [
+        "landscape",  # layout selection
+        "a very detailed prompt",  # prompt rewrite
+    ]
+    mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
+    mock_database.get_image_generation_time_estimate.return_value = None
+
+    with patch.object(main_module, "LAYOUT_SIZES", LAYOUT_TEST_SIZES):
+        asyncio.run(main_module.doodlebob(mock_ctx, message="a scene"))
+
+    rewrite_call = mock_llama_wrapper.chat_completion_instruct.call_args_list[1]
+    expected = main_module.IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE.format(
+        layout="landscape"
+    )
+    assert rewrite_call.kwargs["system_prompt"] == expected
+    assert rewrite_call.kwargs["user_prompt"] == "a scene"
+    assert mock_llama_wrapper.image_generation.call_args.kwargs["prompt"] == (
+        "a very detailed prompt"
+    )
+
+
 def test_doodlebob_selects_portrait(
    mock_ctx: MagicMock,
    mock_llama_wrapper: MagicMock,