[No Effect] Does Weight Syntax (element:1.4) Actually Work? — Thorough Comparison Across 5 Categories × 3 Seeds

[No Effect] Does Weight Syntax (element:1.4) Actually Work? — Thorough Comparison Across 5 Categories × 3 Seeds

Conclusions

In z-image-turbo, no change in attribute strength was confirmed from weight syntax like (element:1.4) across all 5 categories.

For expression, composition, lighting, style, and subject attributes, no visual difference was observed between the unweighted (1.0 equivalent) and (element:1.4) conditions.

However, even with a fixed seed, adding parentheses, a colon, and a number changes the token sequence, which does cause the overall image to change. This is not an effect of the weight value — it’s a side effect of the changed token sequence.

Experiment Conditions

ItemValue
Modelz-image-turbo
Steps8
Samplereuler
Schedulerddim_uniform
CFG1.0
Image Size1024×1024
Seeds42, 7295072554507705269, 4517457392071889496

For each category, 2 conditions (unweighted vs. (element:1.4)) were compared across 3 seeds (30 images total).

Category A: Expression (smiling)

Conditions

ConditionPrompt
Unweighted1girl, 32yo japanese actress, smiling, standing in park, casual outfit, natural lighting, upper body
Weight 1.41girl, 32yo japanese actress, (smiling:1.4), standing in park, casual outfit, natural lighting, upper body

Results

Seedsmiling(smiling:1.4)
42
7295…
4517…

Observation

Smiling in both conditions across all 3 seeds. No difference in degree of smile. Even with fixed seeds, clothing and composition changed.

Note: Even with a fixed seed, adding parentheses and a colon changes the outfit. This stems from the changed token sequence, not the weight effect itself — easy to mistake for the weighting having “worked.”

Category B: Composition (from below)

Conditions

ConditionPrompt
Unweighted1girl, 32yo japanese actress, standing in park, from below, casual outfit, natural lighting, full body
Weight 1.41girl, 32yo japanese actress, standing in park, (from below:1.4), casual outfit, natural lighting, full body

Results

Seedfrom below(from below:1.4)
42
7295…
4517…

Observation

Low-angle (looking-up) composition in both conditions across all 3 seeds. No difference in angle intensity.

Category C: Lighting (strong backlighting)

Conditions

ConditionPrompt
Unweighted1girl, 32yo japanese actress, standing at beach, casual outfit, strong backlighting, upper body
Weight 1.41girl, 32yo japanese actress, standing at beach, casual outfit, (strong backlighting:1.4), upper body

Results

Seedstrong backlighting(strong backlighting:1.4)
42
7295…
4517…

Observation

Both conditions have a backlit atmosphere across all 3 seeds. No difference in backlight intensity. Composition and clothing changed even with fixed seeds.

Category D: Style (film grain)

Conditions

ConditionPrompt
Unweighted1girl, 32yo japanese actress, standing at izakaya entrance, casual outfit, natural lighting, film grain, upper body
Weight 1.41girl, 32yo japanese actress, standing at izakaya entrance, casual outfit, natural lighting, (film grain:1.4), upper body

Results

Seedfilm grain(film grain:1.4)
42
7295…
4517…

Observation

Both conditions show a film-like tone. No difference in grain intensity.

Category E: Subject Attributes (freckles)

Conditions

ConditionPrompt
Unweighted1girl, 32yo japanese actress, freckles, standing in park, casual outfit, natural lighting, upper body, close-up portrait
Weight 1.41girl, 32yo japanese actress, (freckles:1.4), standing in park, casual outfit, natural lighting, upper body, close-up portrait

Results

Seedfreckles(freckles:1.4)
42
7295…
4517…

Observation

No difference in freckle frequency or density.

Summary

CategoryParameterDifference 1.0 vs 1.4
ExpressionsmilingNo difference (3/3 seeds)
Compositionfrom belowNo difference (3/3 seeds)
Lightingstrong backlightingNo difference (3/3 seeds)
Stylefilm grainNo difference (3/3 seeds)
Subject AttributesfrecklesNo difference (3/3 seeds)

z-image-turbo is a distilled model operating at CFG=1.0, and weight syntax (element:weight) does not function for controlling attribute strength. Rather than spending time fine-tuning weight values, it is more effective to optimize prompts through element selection and word order.

Note: No difference was confirmed across all 5 categories — a definitive result. Element selection and word order, rather than weight syntax, form the basic prompt strategy for z-image-turbo.

⚠ 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-best-practices

⚠ 関連記事が見つかりません: /en/tips/breast-size-weight-syntax-test