Seed Variation Baseline — What Changes and What Doesn't with the Same Prompt

Conclusions

Attributes Controllable via Prompt (Stable Against Seed Variation)

Attributes Randomized Per Seed (Hard to Control)

Key Findings

1. “Specified attributes are stable; unspecified attributes are randomized”

This is the clearest pattern. Specifying beige oversized knit sweater reproduces it in 9/9, but if you don’t specify clothing in a simple prompt, all 9 images have different outfits.

2. white background is ignored by the model

All 9 images produced gray-to-brown wall + floor backgrounds. z-image-turbo strongly tends to output dataset-style backgrounds for standing person photos — white background is not an effective instruction.

3. Faces cannot be controlled

A different person is generated every time. The broad category of “young Japanese woman” is stable, but individual facial features depend entirely on the seed.

4. Implications for Future Experiments

When comparing two prompts without fixed seeds, the following attribute differences are within seed variation range and cannot be attributed to prompt changes:

• Differences in facial features
• Differences in clothing patterns
• Differences in specific background content
• Minor hair length differences (bob to semi-long)
• Differences in hand position (hand on cheek vs lap)
• Minor overall color tone differences (warm to cool neutral)

Conversely, the following attribute changes are more likely to be effects of prompt changes:

• Composition (full body vs bust-up, etc.)
• Body orientation or gaze direction
• Type of expression
• Lighting direction or quality
• Changes in clothing color or shape
• Presence or absence of specific objects

Purpose of This Experiment

In AI image generation, different seeds (random number seeds) produce different images even with the same prompt. So when you see a difference after changing a prompt element — is it a prompt change effect or just natural seed variation?

Without this baseline, claims like “removing this element changed the image” don’t hold up. This experiment generates 9 images with the same prompt but different seeds and observes what’s stable and what varies.

Experiment Conditions

Prompts Used

3 prompts were prepared, with scene description specificity varying incrementally.

A: Simple Prompt Results

Observations

Simple Prompt Characteristics

• Specified attributes (subject, age, posture, expression) reproduced stably
• Unspecified attributes (clothing) completely randomized per seed
• white background was ignored by the model — all 9 images show a studio wall + floor background. The model strongly tends toward this type of background for standing person photos
• Faces are not the same person but are similar (rounder face, larger eyes)

B: Café Snapshot Prompt Results

Observations

Café Prompt Characteristics

• Scene description tags powerfully lock in composition — all 9 images consistently match “sitting at window, looking outside” composition and angle
• Sweater color and shape perfectly reproduced — clothing is stable when specified concretely
• On the other hand, unspecified details (café interior, outside view, accessories) are completely different every time
• Faces are different people every time, but within the “young Japanese woman” range

Note: beige oversized knit sweater appearing in all 9 out of 9 images underscores how much prompt specificity matters.

C: Summer Festival Polaroid Prompt Results

Observations

Summer Festival Prompt Characteristics

• Style keyword polaroid photo reproduced frame in 8/9 — high stability but not 100%
• holding cotton candy reproduced in 9/9 — object holding specifications are very stable
• Yukata base color (white to cream) is stable, but pattern and obi change each time — yukata alone doesn’t control the pattern
• 1 image produced a square film style (no frame) — polaroid interpreted as “film photo in general”

Note: The rule “specified attributes are stable, unspecified attributes are random” came through clearly. Write attributes explicitly if you don’t want variation across a batch; leave them unspecified if you want variety. This distinction is fundamental for future experiments.

Related Articles

[Verified] Image Generation Prompt Best Practices

Tag List vs Natural Language — Does Prompt Format Change the Output?