AI Technical Artist / AI image generation Dataset Generation (750 Images)

I am building a private visual-preference research system. I need a dataset of 250 distinct fictional people, with each identity shown in the exact same controlled capture protocol.
This is not for public stock photography, glamour content, or trait classification. The output will be used for private ML pipeline QA, silhouette extraction.
Required deliverables
For each of 250 fictional people, provide exactly these three linked images:
1. HEAD_FRONT — neutral frontal head-and-neck capture.
2. BODY_FRONT — neutral frontal full-body capture.
3. BODY_SIDE — strict 90-degree left-side full-body capture.
That means the core delivery is 750 images total, not 250 images.
The same fictional person must remain visually consistent across all three views: same face, hair, skin tone, age appearance, body proportions, clothing, lighting, camera, and background.
Do not create the side view from a separate text prompt. Use a reliable identity-locking/multi-view workflow—preferably shared 3D geometry, a character-consistency workflow, or source-preserving image editing.
Required capture contract
Every image must show exactly one fictional person, clearly aged 21-23.

• Photorealistic, natural, studio capture.
• Neutral closed-mouth facial expression.
• Straight posture; no glamour pose, twisting, arching, flexing, hip shift, stomach pulling, or exaggerated pose.
• Arms slightly away from torso.
• Plain matte warm-gray studio background.
• Same diffuse lighting and camera height/distance for every identity.
• No filters, text, watermark, logos, props, jewelry, glasses, hats, tattoos, visible branding, or patterned clothing.
• Hair must remain consistent within each person's three images.
• Full body images must include head, hands, legs, and feet without cropping.
• Side image must be a true 90-degree profile: one eye, one ear, one arm, and no visible far shoulder. No three-quarter views.

Clothing
For both body images, every identity must wear the same type of clothing:

• Plain matte aqua dark blue fitted athletic top, full-length and covering the abdomen and waistband.
• Plain matte aqua dark blue opaque fitted leggings.
• Clothing must be correctly sized, non-padded, non-push-up, non-shaping, and without heavy compression.
• No crop tops, exposed midriff, loose layers, shapewear, logos, or patterns.

The purpose is to capture visible clothing-covered outer silhouette and proportions—not inferred anatomy.
Required variation
All identities must be , but the set must deliberately include natural, anatomically plausible variation. Do not make everyone the same slim “model” body type.
Create a balanced, documented coverage plan across these visual silhouette controls:

• Overall frame: 5 balanced groups.
• Clothing-covered upper-torso outer contour: 5 balanced groups, from subtle to fuller.
• Waist-to-hip outer contour: 5 balanced groups.
• Hip/lower-body outer contour: 5 balanced groups.
• Broad variety in facial appearance, hair, skin tone, and adult appearance.

Each of the five groups should contain approximately 50 identities for each dimension. Do not use implausible bodies.
These variation labels are for generation QA only. They will not be used as model features or shown as explanations to users.
Image quality

• Native high-resolution output only: minimum 1536 × 2304 for body images and 1536 × 1536 for head images.
• PNG preferred; no JPEG artifacts, grain, fake upscaling, blur, or compression.
• Realistic skin, hands, feet, face symmetry, and clothing anatomy.
• No duplicated persons, near-duplicate faces, repeated bodies, or copied backgrounds.

Required metadata and files
Deliver a ZIP/folder structure like:
person-beta-001/
head_front.png
body_front.png
body_side.png
manifest.json
Each manifest.json must include:
{
"identityId": "female-beta-001",
"fictionalAdult": true,
"genderScope": "woman",
"views": ["HEAD_FRONT", "BODY_FRONT", "BODY_SIDE"],
"generationMethod": "describe exact workflow",
"modelAndVersion": "exact model/version",
"seedOrSourceReference": "reproducibility information",
"identityLockMethod": "how consistency was preserved",
"generationControls": {
"overallFrame": "generation QA label",
"upperTorsoOuterContour": "generation QA label",
"waistHipOuterContour": "generation QA label",
"lowerBodyOuterContour": "generation QA label"
},
"commercialResearchRightsConfirmed": true
}
Also provide:

• One master CSV/JSON manifest for all 250 identities.
• SHA-256 hashes for every delivered image.
• Prompts, seeds, workflow settings, model/version, and source/edit references needed to reproduce each image.
• Confirmation that every person is entirely fictional, adult, original, and not based on a real person, celebrity, influencer, or copyrighted reference image.
• Full commercial/private ML research rights transferred to me.

Extra nuisance-robustness deliverable
For 30 randomly selected approved identities, provide three extra variants each:

• Minor lighting variation only.
• Minor background-tone variation only.
• Different rendering seed only.

The person's face, body geometry, pose, clothing, and camera framing must remain unchanged.
This is 90 additional images. Quote this separately from the 750-image core set.
Milestones and acceptance
Do not begin the full 250-identity batch immediately.
1. 3 complete identities / 9 images.
2. Review and approval.
3. Paid milestone: 10 complete identities / 30 images.
4. Review and approval.
5. Full remaining batch only after both earlier milestones pass.
I will reject and require correction for:

• Any male, ambiguous-age, or non-adult-looking subject.
• Any independent/unlinked front and side identity.
• Three-quarter side images.
• Cropped feet/hands/head.
• Expression, pose, clothing, background, or lighting violations.
• Near duplicates.
• Grainy, low-detail, upscaled, or visibly AI-broken images.
• Missing metadata or insufficient commercial/private ML rights.

What to include in your proposal
Please show:
1. A previous example of the same fictional person rendered consistently in front, head, and strict 90-degree side views.
2. Your exact identity-locking workflow.
3. Your actual native output resolution.
4. Whether you can provide reproducibility metadata and commercial research rights.
5. Fixed price for the 3-identity test, 10-identity test, 250-identity batch, and 90 nuisance variants.
6. Estimated turnaround after each approval milestone.