2nd/10_Wiki/Topics/Domain_Programming/AI_and_ML/Prompt Weight.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-prompt-weight	Prompt Weight	10_Wiki/Topics	verified	self	Prompt Weighting Attention Weighting Token Emphasis Prompt Strength	none	A	0.9	applied	prompt-engineering generative-ai stable-diffusion midjourney image-gen		2026-05-10	pending	language framework Python diffusers / ComfyUI / Automatic1111 / Midjourney

Prompt Weight

매 한 줄

"매 emphasize / de-emphasize specific tokens in a prompt — (word:1.3) syntax of Stable Diffusion, --w of Midjourney, attention scaling under the hood". AUTOMATIC1111 (2022) 의 prompt-weight syntax 가 community standard 로 자리잡음. 2026 currently FLUX, SD3.5, SDXL Turbo, Midjourney v7 모두 weighting 지원; T5-encoded models 는 syntax 가 다름.

매 핵심

매 syntax (Stable Diffusion / A1111 / ComfyUI)

• (word) — weight ×1.1.
• ((word)) — ×1.21.
• (word:1.3) — explicit weight ×1.3.
• [word] — weight ÷1.1.
• [word:0.5] — weight ×0.5.
• (red hair:1.4) (blue eyes:0.8) — phrase-level.

매 syntax (Midjourney v7)

• cat dog — equal weight.
• cat::2 dog::1 — double-colon multi-prompt with weights.
• --w 0.5 — image weight (text vs reference image).
• --s 250 — stylize strength.

매 syntax (FLUX / T5-encoded)

• T5 understands natural language; (word:1.3) syntax 매 mostly ignored.
• Use emphasis via wording: "very prominent X", "subtle hint of Y".
• Some forks (forge, ComfyUI) 매 still parse weights via re-prompting.

매 mechanism (under the hood)

• CLIP/T5 text encoder → token embeddings.
• A1111: weight w → multiply token embedding by w (post-encoding rescale).
• Compel library: more sophisticated — interpolates between conditioning vectors.
• Cross-attention scaling: alternative — scale K/V at attention layer.

매 best practices

• Stay between 0.5 and 1.5; 매 above 1.5 → distortion / saturation.
• Negative prompts often more effective than [word] syntax.
• Long prompts: weight the critical 3-5 tokens, leave rest at 1.0.
• For T5 models, use natural-language emphasis instead.

💻 패턴

diffusers + Compel (programmatic weighting)

from diffusers import StableDiffusionXLPipeline
from compel import Compel, ReturnedEmbeddingsType
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")

compel = Compel(
    tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
    text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
    returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
    requires_pooled=[False, True],
)

prompt = "a (red:1.4) sports car on a (sunny:0.7) beach, cinematic"
conditioning, pooled = compel(prompt)
image = pipe(prompt_embeds=conditioning, pooled_prompt_embeds=pooled).images[0]

A1111-style parsing (manual)

import re
def parse_weighted(prompt):
    """Return list of (text, weight) tuples."""
    out, depth_paren, depth_brack = [], 0, 0
    # Simplified: handles (text:1.3) only
    pattern = re.compile(r"\(([^():]+):([\d.]+)\)")
    parts, last = [], 0
    for m in pattern.finditer(prompt):
        if m.start() > last:
            parts.append((prompt[last:m.start()], 1.0))
        parts.append((m.group(1), float(m.group(2))))
        last = m.end()
    if last < len(prompt):
        parts.append((prompt[last:], 1.0))
    return parts

Cross-attention scaling (Hugging Face)

# Scale a specific token's attention by factor
from diffusers.models.attention_processor import AttnProcessor

class WeightedAttn(AttnProcessor):
    def __init__(self, token_idx, scale):
        self.token_idx, self.scale = token_idx, scale
    def __call__(self, attn, hidden, encoder_hidden, attention_mask=None):
        # In encoder_hidden, multiply token_idx slot by scale before attn
        encoder_hidden = encoder_hidden.clone()
        encoder_hidden[:, self.token_idx] *= self.scale
        return super().__call__(attn, hidden, encoder_hidden, attention_mask)

Midjourney prompt

masterpiece anime girl::3 cyberpunk city background::1 neon lights::0.5
--ar 16:9 --s 500 --v 7

FLUX-style natural-language emphasis (no syntax)

# Bad (FLUX ignores):  "(red hair:1.5) girl"
# Good:                "girl with strikingly vivid red hair, the red is the most prominent color in the image"

Prompt-blending (interpolate two prompts)

from compel import Compel
c1 = compel("a cat in a forest")
c2 = compel("a robot in a city")
mixed = (c1 + c2) / 2  # Compel supports tensor arithmetic
image = pipe(prompt_embeds=mixed).images[0]

Step-conditional weighting ([from:to:step])

[cat:dog:0.5] in a field
# 0-50% steps: "cat", 50-100%: "dog"
# Useful for changing subject mid-denoising

매 결정 기준

상황	Approach
SDXL / SD1.5 / SD2.1	A1111 (word:1.3) syntax via Compel
FLUX / SD3.5 (T5)	Natural-language emphasis
Midjourney v7	::weight syntax
Subject + style mix	Multi-prompt with :: or compel blends
Subtle adjustment	0.8-1.2 range
Strong push	1.3-1.5; rarely above
Suppress concept	Negative prompt (preferred) over [word]

기본값: Compel for SDXL programmatic; A1111 syntax for casual; natural language for FLUX.

🔗 Graph

• 부모: Prompt_Engineering · AI 이미지 생성 (AI Image Generation)
• 변형: Negative Prompt
• 응용: Stable-Diffusion · FLUX · Midjourney · ComfyUI
• Adjacent: CLIP · Diffusion-Models · ControlNet

🤖 LLM 활용

언제: image generation pipelines, fine-grained subject/style control, automated prompt synthesis. 언제 X: text-only LLM prompts (GPT/Claude don't use this syntax — use emphasis words instead), T5-only models.

❌ 안티패턴

• Weight > 2.0: 매 saturated artifacts, deformed output.
• Stacking parens (((((word))))): hard to read; use explicit (word:1.6).
• A1111 syntax on FLUX/T5: silently ignored — switch to natural language.
• Weighting every token: dilutes effect; pick 2-4 priorities.
• Forgetting negative prompt: often the right tool for "not X".

🧪 검증 / 중복

• Verified (AUTOMATIC1111 wiki, Compel docs, Midjourney v7 docs 2024-2025, FLUX official guidance).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — canonical prompt-weight ref + FLUX/T5 caveat