Add natural language formatting for T5-XXL
This commit builds on PR #1 (independence + debug fixes) and adds natural language formatting to improve T5-XXL prompt quality. Changes: - Added strip_weights_for_natural_language() method to remove weights from T5-XXL - Added _build_natural_language_sentence() helper for consistent templates - Integrated photo_type into opening sentences (e.g., 'A birds eye view photography of...') - Converted roles and hairstyles to natural language sentences - Converted clothing to sentence format ('Dressed in...') - Converted composition and pose to complete sentences - Grouped physical features into three coherent sentences: - Face, eyes, and expression ('They have...') - Skin, age, and ethnicity ('The subject has...') - Hair, accessories, and markings ('Notable features include...') - Added device description at end ('The image was captured using a...') - Fixed photo_type CLIP_L re-fetch bug by reusing same value from T5-XXL - Enhanced clean_prompt_string() with better period and spacing handling - Updated RETURN_TYPES to return 6 outputs: prompt, t5xxl, clip_l, clip_g, seed, debug_info Benefits: - T5-XXL receives natural language matching its training - CLIP_L and CLIP_G retain keyword format with weights - More consistent results (no RNG re-fetch issues) - Better readability and debugging
This commit is contained in:
John B
+192
-42
@@ -144,17 +144,61 @@ class PromptGenerator:
|
||||
text = re.sub(r'\s+(BREAK_CLIP[GL])', r' \1', text)
|
||||
# Remove spaces after BREAK markers
|
||||
text = re.sub(r'(BREAK_CLIP[GL])\s+', r'\1 ', text)
|
||||
# Remove spaces before periods
|
||||
text = re.sub(r'\s+\.', '.', text)
|
||||
# Ensure space after period before word (but not multiple spaces)
|
||||
text = re.sub(r'\.([A-Z])', r'. \1', text)
|
||||
# Fix double periods
|
||||
text = re.sub(r'\.\.+', '.', text)
|
||||
# Replace multiple spaces with a single space
|
||||
text = re.sub(r'\s+', ' ', text).strip()
|
||||
# Specific cleanup like 'of a as' -> 'of a' etc.
|
||||
text = text.replace(" of as ", " of ") # Check variations if needed
|
||||
text = text.replace(" a as ", " as ") # Example
|
||||
text = text.replace(" of as ", " of ")
|
||||
text = text.replace(" a as ", " as ")
|
||||
text = text.replace(" as as ", " as ") # Fix "working as as a"
|
||||
text = text.replace(" with with ", " with ") # Fix "with with"
|
||||
# Remove empty segments like ", ," or leading ", " after splits
|
||||
text = re.sub(r'^,\s*', '', text)
|
||||
text = re.sub(r'\s*,\s*$', '', text)
|
||||
text = re.sub(r',(\s*,)+', ',', text)
|
||||
return text
|
||||
|
||||
def strip_weights_for_natural_language(self, text):
|
||||
"""Removes parentheses and weights from text for natural language (T5-XXL)."""
|
||||
if not text: return ""
|
||||
# Remove weights like (text:1.5) -> text
|
||||
text = re.sub(r'\(([^:)]+):[0-9.]+\)', r'\1', text)
|
||||
# Remove simple parentheses (text) -> text
|
||||
text = re.sub(r'\(([^)]+)\)', r'\1', text)
|
||||
return text
|
||||
|
||||
def _build_natural_language_sentence(self, category, value, context=None):
|
||||
"""Converts a category value into a natural language fragment."""
|
||||
if not value:
|
||||
return ""
|
||||
|
||||
templates = {
|
||||
'role': f"working as {value}",
|
||||
'hairstyle': f"with {value}",
|
||||
'clothing': f"dressed in {value}",
|
||||
'composition': f"The composition follows {value}",
|
||||
'pose': f"The subject is {value}",
|
||||
'face_features': f"with {value}",
|
||||
'eye_color': f"with {value} eyes",
|
||||
'skin_tone': f"with {value} skin",
|
||||
'age_group': f"appearing {value}",
|
||||
'ethnicity': f"of {value} descent",
|
||||
'accessories': f"wearing {value}",
|
||||
'expression': f"with a {value} expression",
|
||||
'tattoos_scars': f"featuring {value}",
|
||||
'hair_color': f"with {value} hair",
|
||||
'body_markings': f"displaying {value}",
|
||||
'facial_hair': f"with {value}",
|
||||
'makeup': f"styled with {value} makeup"
|
||||
}
|
||||
|
||||
return templates.get(category, value)
|
||||
|
||||
def _format_debug_info(self, debug_info):
|
||||
"""Format debug info as readable string"""
|
||||
lines = []
|
||||
@@ -197,6 +241,10 @@ class PromptGenerator:
|
||||
# Original prompt is T5 content with markers removed
|
||||
original_content = t5xxl_content.replace('BREAK_CLIPL', '').replace('BREAK_CLIPG', '')
|
||||
|
||||
# Strip weights/parentheses from T5-XXL content (natural language doesn't need weights)
|
||||
original_content = self.strip_weights_for_natural_language(original_content)
|
||||
t5xxl_content = self.strip_weights_for_natural_language(t5xxl_content)
|
||||
|
||||
# Clean all parts
|
||||
original_clean = self.clean_prompt_string(original_content)
|
||||
t5xxl_clean = self.clean_prompt_string(t5xxl_content)
|
||||
@@ -223,36 +271,55 @@ class PromptGenerator:
|
||||
if custom:
|
||||
components.append(custom)
|
||||
|
||||
# --- 2. Artform / Style Lead-in ---
|
||||
# --- 2. Artform / Style Lead-in (with photo_type integrated) ---
|
||||
artform = self._get_choice(kwargs.get("artform", "disabled"), ARTFORM)
|
||||
is_photographer = (artform.lower() == "photography")
|
||||
|
||||
# Get photo_type early to integrate into opening
|
||||
photo_type = self._get_choice(kwargs.get("photo_type", "random"), PHOTO_TYPE)
|
||||
|
||||
if is_photographer:
|
||||
photo_style = self._get_choice(kwargs.get("photography_styles", "random"), PHOTOGRAPHY_STYLES)
|
||||
if photo_style:
|
||||
components.append(photo_style)
|
||||
else:
|
||||
# Fallback if random selects nothing or style is disabled
|
||||
components.append("photography") # Default base style
|
||||
# Build opening with optional photo_type
|
||||
opening_parts = []
|
||||
if photo_type:
|
||||
opening_parts.append(f"A {photo_type}")
|
||||
opening_parts.append(photo_style if photo_style else "photography")
|
||||
components.append(" ".join(opening_parts))
|
||||
|
||||
# Add "of" if a subject or default tag will follow
|
||||
if kwargs.get("subject") or kwargs.get("default_tags", "disabled").lower() != "disabled":
|
||||
components.append("of") # Add connector word
|
||||
|
||||
elif artform and artform.lower() != "disabled":
|
||||
components.append(artform)
|
||||
# Build opening with optional photo_type
|
||||
opening_parts = []
|
||||
if photo_type:
|
||||
opening_parts.append(f"A {photo_type}")
|
||||
opening_parts.append(artform)
|
||||
components.append(" ".join(opening_parts))
|
||||
|
||||
# Add "of" if a subject or default tag will follow and artform isn't inherently descriptive like 'illustration'
|
||||
if kwargs.get("subject") or kwargs.get("default_tags", "disabled").lower() != "disabled":
|
||||
# Could refine this list if needed
|
||||
if artform.lower() not in ["illustration", "painting", "drawing", "sketch"]:
|
||||
components.append("of")
|
||||
|
||||
elif photo_type:
|
||||
# Standalone photo_type when artform is disabled
|
||||
components.append(f"A {photo_type}")
|
||||
# Add "of" if a subject or default tag will follow
|
||||
if kwargs.get("subject") or kwargs.get("default_tags", "disabled").lower() != "disabled":
|
||||
components.append("of")
|
||||
|
||||
# --- 3. Subject Definition ---
|
||||
subject = kwargs.get("subject", "")
|
||||
default_tags_input = kwargs.get("default_tags", "random")
|
||||
body_type_input = kwargs.get("body_types", "random")
|
||||
|
||||
chosen_subject_elements = []
|
||||
if subject: # User provided subject takes precedence
|
||||
# User provided subject takes precedence (but not if it's "random" or "disabled")
|
||||
if subject and subject.lower() not in ["random", "disabled"]:
|
||||
chosen_body_type = self._get_choice(body_type_input, BODY_TYPES)
|
||||
if chosen_body_type:
|
||||
chosen_subject_elements.extend(["a", chosen_body_type]) # e.g., "a muscular"
|
||||
@@ -282,24 +349,41 @@ class PromptGenerator:
|
||||
resolved_subject_desc = " ".join(chosen_subject_elements).lower()
|
||||
|
||||
# --- 4. Core Details (Roles, Hairstyles, Additional Details) ---
|
||||
core_details = [
|
||||
self._get_choice(kwargs.get("roles", "random"), ROLES),
|
||||
self._get_choice(kwargs.get("hairstyles", "random"), HAIRSTYLES),
|
||||
self._get_choice(kwargs.get("additional_details", "random"), ADDITIONAL_DETAILS),
|
||||
]
|
||||
components.append(smart_join(core_details))
|
||||
# Build natural language sentences instead of comma-joining
|
||||
role = self._get_choice(kwargs.get("roles", "random"), ROLES)
|
||||
hairstyle = self._get_choice(kwargs.get("hairstyles", "random"), HAIRSTYLES)
|
||||
additional_details = self._get_choice(kwargs.get("additional_details", "random"), ADDITIONAL_DETAILS)
|
||||
|
||||
core_parts = []
|
||||
if role:
|
||||
core_parts.append(self._build_natural_language_sentence('role', role))
|
||||
if hairstyle:
|
||||
core_parts.append(self._build_natural_language_sentence('hairstyle', hairstyle))
|
||||
|
||||
if core_parts:
|
||||
components.append(" ".join(core_parts))
|
||||
|
||||
if additional_details:
|
||||
components.append(f". {additional_details}")
|
||||
|
||||
# --- 5. Clothing ---
|
||||
clothing = self._get_choice(kwargs.get("clothing", "random"), CLOTHING)
|
||||
if clothing:
|
||||
components.append(f"dressed in {clothing}")
|
||||
components.append(f". Dressed in {clothing}")
|
||||
|
||||
# --- 6. Composition & Pose ---
|
||||
comp_pose = [
|
||||
self._get_choice(kwargs.get("composition", "random"), COMPOSITION),
|
||||
self._get_choice(kwargs.get("pose", "random"), POSE),
|
||||
]
|
||||
components.append(smart_join(comp_pose))
|
||||
# Build natural language sentences with periods
|
||||
composition = self._get_choice(kwargs.get("composition", "random"), COMPOSITION)
|
||||
pose = self._get_choice(kwargs.get("pose", "random"), POSE)
|
||||
|
||||
comp_pose_parts = []
|
||||
if pose:
|
||||
comp_pose_parts.append(f"The subject is {pose}")
|
||||
if composition:
|
||||
comp_pose_parts.append(f"The composition follows {composition}")
|
||||
|
||||
if comp_pose_parts:
|
||||
components.append(". " + ". ".join(comp_pose_parts))
|
||||
|
||||
# --- BREAK CLIP G 1 ---
|
||||
components.append("BREAK_CLIPG")
|
||||
@@ -328,48 +412,107 @@ class PromptGenerator:
|
||||
components.append("BREAK_CLIPG")
|
||||
|
||||
# --- 9. Physical Features ---
|
||||
features = [
|
||||
self._get_choice(kwargs.get("face_features", "random"), FACE_FEATURES),
|
||||
self._get_choice(kwargs.get("eye_colors", "random"), EYE_COLORS),
|
||||
self._get_choice(kwargs.get("skin_tone", "random"), SKIN_TONE),
|
||||
self._get_choice(kwargs.get("age_group", "random"), AGE_GROUP),
|
||||
self._get_choice(kwargs.get("ethnicity", "random"), ETHNICITY),
|
||||
self._get_choice(kwargs.get("accessories", "random"), ACCESSORIES),
|
||||
self._get_choice(kwargs.get("expression", "random"), EXPRESSION),
|
||||
self._get_choice(kwargs.get("tattoos_scars", "random"), TATTOOS_SCARS),
|
||||
self._get_choice(kwargs.get("hair_color", "random"), HAIR_COLOR),
|
||||
self._get_choice(kwargs.get("body_markings", "random"), BODY_MARKINGS),
|
||||
self._get_choice(kwargs.get("facial_hair", "random"), FACIAL_HAIR),
|
||||
self._get_choice(kwargs.get("makeup_styles", "random"), MAKEUP_STYLES),
|
||||
]
|
||||
# Get all feature values
|
||||
face_features = self._get_choice(kwargs.get("face_features", "random"), FACE_FEATURES)
|
||||
eye_color = self._get_choice(kwargs.get("eye_colors", "random"), EYE_COLORS)
|
||||
skin_tone = self._get_choice(kwargs.get("skin_tone", "random"), SKIN_TONE)
|
||||
age_group = self._get_choice(kwargs.get("age_group", "random"), AGE_GROUP)
|
||||
ethnicity = self._get_choice(kwargs.get("ethnicity", "random"), ETHNICITY)
|
||||
accessories = self._get_choice(kwargs.get("accessories", "random"), ACCESSORIES)
|
||||
expression = self._get_choice(kwargs.get("expression", "random"), EXPRESSION)
|
||||
tattoos_scars = self._get_choice(kwargs.get("tattoos_scars", "random"), TATTOOS_SCARS)
|
||||
hair_color = self._get_choice(kwargs.get("hair_color", "random"), HAIR_COLOR)
|
||||
body_markings = self._get_choice(kwargs.get("body_markings", "random"), BODY_MARKINGS)
|
||||
|
||||
components.append(smart_join(features))
|
||||
# Facial hair and makeup - now independent of gender for modern/creative contexts
|
||||
facial_hair = self._get_choice(kwargs.get("facial_hair", "random"), FACIAL_HAIR)
|
||||
makeup = self._get_choice(kwargs.get("makeup_styles", "random"), MAKEUP_STYLES)
|
||||
|
||||
# Group features into coherent sentences
|
||||
feature_sentences = []
|
||||
|
||||
# Sentence 1: Face, eyes, expression
|
||||
face_parts = []
|
||||
if face_features:
|
||||
face_parts.append(face_features)
|
||||
if eye_color:
|
||||
face_parts.append(f"{eye_color} eyes")
|
||||
if expression:
|
||||
face_parts.append(f"a {expression} expression")
|
||||
if face_parts:
|
||||
feature_sentences.append("They have " + " and ".join(face_parts))
|
||||
|
||||
# Sentence 2: Skin, age, ethnicity
|
||||
identity_parts = []
|
||||
if skin_tone:
|
||||
identity_parts.append(f"{skin_tone} skin")
|
||||
if age_group:
|
||||
# Age groups that are nouns need articles
|
||||
age_nouns = ["adult", "child", "infant", "preteen", "senior", "teenager", "toddler", "young adult"]
|
||||
if age_group in age_nouns:
|
||||
article = "an" if age_group[0].lower() in "aeiou" else "a"
|
||||
identity_parts.append(f"appearing as {article} {age_group}")
|
||||
else:
|
||||
# Adjectives don't need articles
|
||||
identity_parts.append(f"appearing {age_group}")
|
||||
if ethnicity:
|
||||
identity_parts.append(f"of {ethnicity} descent")
|
||||
if identity_parts:
|
||||
feature_sentences.append("The subject has " + ", ".join(identity_parts))
|
||||
|
||||
# Sentence 3: Hair, accessories, markings, gender-specific (only one of facial_hair OR makeup)
|
||||
detail_parts = []
|
||||
if hair_color:
|
||||
detail_parts.append(f"{hair_color} hair")
|
||||
if accessories:
|
||||
detail_parts.append(f"wearing {accessories}")
|
||||
if tattoos_scars:
|
||||
detail_parts.append(f"featuring {tattoos_scars}")
|
||||
if body_markings:
|
||||
detail_parts.append(f"displaying {body_markings}")
|
||||
if facial_hair and not makeup: # Only add if makeup isn't present
|
||||
detail_parts.append(f"{facial_hair}")
|
||||
elif makeup: # Only add if facial_hair isn't present
|
||||
detail_parts.append(f"{makeup}")
|
||||
if detail_parts:
|
||||
feature_sentences.append("Notable features include " + ", ".join(detail_parts))
|
||||
|
||||
if feature_sentences:
|
||||
components.append(". " + ". ".join(feature_sentences))
|
||||
|
||||
# --- 10. Camera/Device (for T5-XXL natural language) ---
|
||||
device = self._get_choice(kwargs.get("device", "random"), DEVICE)
|
||||
if device:
|
||||
components.append(f". The image was captured using a {device}")
|
||||
|
||||
# --- BREAK CLIP L 1 ---
|
||||
components.append("BREAK_CLIPL")
|
||||
|
||||
# --- 10. Technical/Artistic Details ---
|
||||
# --- 11. Technical/Artistic Details (for CLIP_L keywords) ---
|
||||
# All categories now work independently - users control via their selections
|
||||
tech_artist_details = []
|
||||
|
||||
photo_type = self._get_choice(kwargs.get("photo_type", "random"), PHOTO_TYPE)
|
||||
# Framing (weighted) - reuse the same value fetched for T5-XXL opening
|
||||
# This ensures consistency between T5-XXL and CLIP_L
|
||||
if photo_type:
|
||||
random_value = round(self.rng.uniform(1.1, 1.5), 1)
|
||||
tech_artist_details.append(f"({photo_type}:{random_value})")
|
||||
|
||||
device = self._get_choice(kwargs.get("device", "random"), DEVICE)
|
||||
# Device - already fetched above for T5-XXL, reuse the same value
|
||||
if device:
|
||||
tech_artist_details.append(f"shot on {device}")
|
||||
|
||||
# Digital artform - works independently
|
||||
digital_artform = self._get_choice(kwargs.get("digital_artform", "random"), DIGITAL_ARTFORM)
|
||||
if digital_artform:
|
||||
tech_artist_details.append(digital_artform)
|
||||
|
||||
# Photographer - works independently
|
||||
photographer = self._get_choice(kwargs.get("photographer", "random"), PHOTOGRAPHER)
|
||||
if photographer:
|
||||
tech_artist_details.append(f"photo by {photographer}")
|
||||
|
||||
# Artist - works independently
|
||||
artist = self._get_choice(kwargs.get("artist", "random"), ARTIST)
|
||||
if artist:
|
||||
tech_artist_details.append(f"by {artist}")
|
||||
@@ -416,6 +559,13 @@ class PromptGenerator:
|
||||
debug_info['not_used']['default_tags'] = "ignored (subject takes precedence)"
|
||||
del debug_info['used']['default_tags']
|
||||
|
||||
# photography_styles only used when artform is photography
|
||||
if kwargs.get('photography_styles', '') and kwargs.get('photography_styles', '').lower() not in ['disabled', '']:
|
||||
if artform.lower() != "photography":
|
||||
if 'photography_styles' in debug_info['used']:
|
||||
debug_info['not_used']['photography_styles'] = f"ignored (artform={artform}, not photography)"
|
||||
del debug_info['used']['photography_styles']
|
||||
|
||||
# Format debug output as readable string
|
||||
debug_output = self._format_debug_info(debug_info)
|
||||
|
||||
@@ -467,7 +617,7 @@ class FluxPromptGenerator:
|
||||
|
||||
# Correct RETURN_TYPES and add RETURN_NAMES
|
||||
RETURN_TYPES = ("STRING", "STRING", "STRING", "STRING", "STRING", "STRING")
|
||||
RETURN_NAMES = ("prompt", "t5xxl", "clip_l", "clip_g", "seed_used", "debug_info")
|
||||
RETURN_NAMES = ("prompt", "t5xxl", "clip_l", "clip_g", "seed", "debug_info")
|
||||
|
||||
FUNCTION = "execute"
|
||||
CATEGORY = "Prompt"
|
||||
@@ -477,7 +627,7 @@ class FluxPromptGenerator:
|
||||
seed = kwargs.get('seed', 0) # Extract seed separately if needed elsewhere
|
||||
prompt_generator = PromptGenerator(seed)
|
||||
prompt = prompt_generator.generate_prompt(**kwargs)
|
||||
# Unpack the 6-tuple and return all outputs including debug_info
|
||||
# Unpack the 6-tuple and return all 6 outputs including seed_used
|
||||
original_clean, seed_used, t5xxl_clean, clip_l_clean, clip_g_clean, debug_output = prompt
|
||||
# Concatenate all sections into a single combined prompt
|
||||
combined_prompt = ", ".join(filter(None, [t5xxl_clean, clip_l_clean, clip_g_clean]))
|
||||
|
||||
Reference in New Issue
Block a user