3 "God Prompts" That Never Miss | Ablation-Verified Minimal Versions

Conclusions

• All 3 god prompts fit within 75 tokens — ablation testing removed unnecessary elements, reducing Summer Festival to 44 tokens, Morning Bed to 54 tokens, and Cafe Snap to 25 tokens
• Simple, non-contradictory environments are the key to stability — fewer elements means fewer failures; consistent lighting and location descriptions matter
• Hand management contributes to stability — techniques like holding cotton candy or resting a chin on hands help stabilize hand rendering
• Don’t add unnecessary quality words — coherent anatomy, natural skin texture, and 8K have no effect in z-image-turbo
• Style specification at the front is most effective — declaring the photo style at the start, like A Polaroid instant photo or An intimate close-up portrait, stabilizes composition and atmosphere

Each element’s necessity was verified one by one via ablation testing, and these are the minimal versions with unnecessary elements removed. Even generating 9 images in a row, every single one hits the mark.

Selection Criteria

• The first-draft prompt produced the intended image with no revisions
• Generating 9 images, every one was stable at goal quality
• Unnecessary elements removed via ablation testing

How Token Count Is Measured

Token counts in this article are measured values using the CLIP tokenizer (openai/clip-vit-large-patch14). Word count and token count do not match. For details, see Prompt Basics - How to Count Tokens.

God Prompt 1: Summer Festival Polaroid

Intent: A Polaroid photo of a woman in a yukata holding cotton candy and smiling under festival lanterns.

Minimal Version — 9 Sample Images

Elements Removed by Ablation Testing

See detailed test results here

75 tokens → 44 tokens (31 tokens removed). The Polaroid border, faded colors, lanterns, yukata, and cotton candy are all preserved.

Comparison with original (click to expand)

Original (75 tokens) コピー

A Polaroid instant photo of a woman. 1girl, 20yo japanese woman, outdoor summer festival grounds at dusk, paper lantern warm light, food stalls blurred in background, light blue yukata, cotton candy in hand, looking at camera, big smile. Polaroid instant film look, slightly faded colors, soft vignette, warm nostalgic tint, fixed focus. coherent anatomy.

Original version samples (for reference):

Why It’s Stable

1. A Polaroid instant photo is extremely effective in CLIP — the training data contains large amounts of Polaroid photography, so white borders, faded colors, and vignette are reproduced all at once
2. The components of “summer festival” are simple and unambiguous — paper lantern + yukata + dusk pins down the scene with no ambiguity
3. cotton candy in hand contributes to hand stability — giving the subject something to hold stabilizes finger rendering
4. 44 tokens leaves plenty of room below the 75-token limit — fitting entirely within one chunk means every element gets full effect

God Prompt 2: Morning Bed Intimate Portrait

Intent: An intimate shot of a woman lying in bed with rumpled white sheets, lit by morning light through curtains.

Minimal Version — 9 Sample Images

Elements Removed by Ablation Testing

85 tokens → 54 tokens (31 tokens removed). Even with the trailing quality instructions removed, the opening An intimate close-up portrait sufficiently defines the composition, bokeh, and intimate atmosphere. Now fits within 75 tokens.

Comparison with original (click to expand)

Original (85 tokens) コピー

An intimate close-up portrait of a woman. 1girl, 20yo japanese woman, in bed, white sheets rumpled, morning light through sheer curtains, oversized white shirt, lying on side, chin resting on hands, half-closed eyes, soft smile, warm ambient light. intimate portrait quality, shallow depth of field, soft bokeh background, gentle lighting on face. coherent anatomy, natural skin texture.

Original version samples (for reference):

Why It’s Stable

1. An intimate close-up portrait simultaneously determines composition and atmosphere — “close-up” and “intimate” covered in one efficient phrase
2. The environment is simple with no contradictions — just in bed, white sheets, morning light through curtains
3. chin resting on hands prevents hand rendering failures — the chin-on-hands pose stabilizes finger depiction
4. half-closed eyes controls expression — the half-open eyes stably convey a “just woke up” / “drowsy” mood

God Prompt 3: Cafe Window Snap

Intent: A natural, casual-looking shot — like a friend caught zoning out at a cafe window seat, taken on a smartphone.

Minimal Version — 9 Sample Images

Elements Removed by Ablation Testing

83 tokens → 25 tokens — a massive reduction. The opening natural language sentence, quality keywords, and unnecessary modifiers are all gone. Scene description tags alone reproduce the composition, lighting, and atmosphere.

Comparison with original (click to expand)

Original (83 tokens) コピー

A candid iPhone snapshot of an actress in her everyday life. 1girl, 22yo japanese woman, small café window seat, natural overcast daylight through glass, beige oversized knit sweater, sitting, looking out window, gentle natural expression. The photo feels imperfect and unposed: slightly awkward crop, mild smartphone compression, no cinematic lighting or editorial polish. photorealistic, snapshot aesthetic, natural skin texture, coherent anatomy.

Original version samples (for reference):

Why It’s Stable

1. Scene description tags are specific and unambiguous — small cafe window seat + natural overcast daylight + beige oversized knit sweater pins down the scene
2. looking out window shifts the gaze away from camera — gives the candid, natural-moment nuance
3. actress guides facial direction — steers toward expressive, attractive facial features (test result)
4. 25 tokens leaves a huge margin below the 75-token limit — fits entirely within one chunk, every element gets full effect

Common Patterns Across God Prompts

The 5 Conditions for a God Prompt

1. Simple environment — fewer elements means fewer failures
2. Hand management — give the subject something to hold, or use a pose where hands aren’t visible
3. No contradicting instructions — lighting and location descriptions must be consistent
4. Within 75 tokens — fitting within one chunk means every element gets full effect
5. No unnecessary quality words — coherent anatomy, natural skin texture, 8K are ineffective

Related Articles

⚠ 関連記事が見つかりません: /en/tips/god-prompt-ablation

⚠ 関連記事が見つかりません: /en/tips/coherent-anatomy-test

⚠ 関連記事が見つかりません: /en/tips/prompt-basics

⚠ 関連記事が見つかりません: /en/tips/prompt-refinement