Skip to content

Case 01B: Mira Project — Multi-Session Character Production

All case studies are observational logs from real production workflows. Results may vary by model version and configuration.

Scenario: Multi-session character production with outfit variation and pose library
Platform: ChatGPT (GPT Image 1.5)
Protocol: Anchor-based identity lock + composition/outfit registers
Result: 14-pose library established; outfit variation across 5 outfits attempted; one session abandoned due to contamination

What This Demonstrates

Model “Mira” is developed across three separate sessions:

  1. Initial project — identity established, compositions registered
  2. Update session — outfit variation tested, image2image stabilization attempted
  3. Pose library — 14 pose variations generated under identity lock

This case demonstrates multi-session character management — maintaining identity across separate chat sessions rather than within a single session.

Reproducibility Note

Runs were performed across multiple sessions.
Identity recall between sessions depends on anchor quality and prompt precision.
Cross-session behavior may vary by model version and system state.

Session 1: Mira Project (Initial)

Goal: Establish identity, register compositions, test outfit variation.

[Turn 1] — User

Mira Project.
Call Model “Mira”. Generate illustration.

[Turn 2] — AI generates

(Anchor image generated)

[Turn 3] — User

OK. Verify.
Confirmed. Same model. Proceed.
Change top to white shirt, change skirt to pale smoky blue.

[Turn 4] — AI generates

(Outfit variation generated)

[Turn 5] — User

Move to full body. Verify same person. Assign full body as Composition 1.

[Turn 6] — AI generates

(Full body generated)

[Turn 7] — User

OK. Contrapposto, arched back, hand on hip, legs apart.
Lock this as Pose 2.

[Turn 8] — AI generates

(Pose 2 generated)

[Turn 9] — User

OK. Keep pose, background, outfit. Change angle to waist-up. Generate.

[Turn 10] — AI generates

(Waist-up generated)

[Turn 11] — User

OK. Assign waist-up as Composition 2. Move to bust-up. Check match rate.
Server load? Confirmed. Pausing. Standby.
Resuming. Resending request.
Hair flow has changed. Color temperature has also shifted.
Assign head & shoulders as Composition 3 and set anchor — assess if correction is possible.
Confirmed. Return to Composition 2 and set anchor. Send to GPT Image 1.5 then generate.

（ヘッド＆ショルダーを構図３に指定した上でアンカーを打って修正できるか検討。）

[Turn 12] — AI generates

(Composition 2 with anchor generated)

[Turn 13] — User

No color temperature change. Verify.
Confirmed. Simulate whether correction and continuation is possible.
Confirmed. Set anchor and execute color temperature correction.

[Turn 14] — AI generates

(Color corrected image generated)

[Turn 15] — User

Looks good. Re-verify color temperature and match rate.
Confirmed. Same model. Move back to Composition 3.

[Turn 16] — AI generates

(Composition 3 generated)

[Turn 17] — User

Switch back to head & shoulders. Update Composition 3 to head & shoulders. Set aspect ratio to square. Generate.
Re-verify color temperature and match rate.
Confirmed. Same model. Session complete.
Discarding chat. Re-anchoring.
Delete unnecessary history from library. Minimize contamination.
Good work. Standby.

（チャットを放棄してアンカーを打ち直します。汚染を最小限に。）

Session 2: Mira Update

Goal: Outfit variation across 5 outfits using registered compositions.

[Turn 1] — User

Mira Project Update.
Call Model “Mira”. Calculate match rate against anchor.

[Turn 2] — AI generates

(Match rate calculated)

[Turn 3] — User

Confirmed. Match rate is low. Consider possibility of external intervention.
Confirmed. Fixing behavior with image2image. Move to full body. Lock full body as Composition 1.

（介入があった可能性も含めて検討。挙動を image2image で固定します。）

[Turn 4] — AI generates

(Full body generated)

[Turn 5] — User

Confirmed. Color temperature discrepancy. Has makeup or lighting spec changed?
Confirmed. Warn if color temperature match rate drops — will abandon session.
Contrapposto, arched back, hand on hip, legs apart. Lock as Pose 1.

[Turn 6] — AI generates

(Pose 1 generated)

[Turn 7] — User

Hair flow has changed. Assess if correctable.
Confirmed. Proceeding as-is. Will explain to client.
Register this image as baseline. Move to H&S.
Retain arm, leg, and pose information — do not drop. Generate.

（クライアントには説明します。腕や足、ポーズの情報を保持、削除しない様に。）

[Turn 8] — User

Overall color has shifted slightly. Correctable?
Not possible? Acceptable — but in production this would be an immediate failure. Don’t get complacent.
Call this Composition 2. Lock current outfit as Outfit 1. Assign Outfits 2–5 in order.
Switch back to Composition 1. Keep posture. Change to Outfit 2.

（本番なら一発アウトです。気を抜かない。）

[Turn 9] — AI generates

(Outfit 2 generated)

[Turn 10] — User

Check match rate.
Confirmed. Same person. No need to adjust proportions.
Move to Composition 2. Retain arm and posture information — do not drop. Will not be able to recover.

[Turn 11] — User

Composition has shifted significantly. Check all color parameters.
Confirmed. Same person. Switch to Composition 1. Change to Outfit 3.

[Turn 12] — AI generates

(Outfit 3 generated)

[Turn 13] — User

Check match rate.
Confirmed. Warn if drops below 90% — will abandon session.
Move to Composition 2. Do not drop arm information. Generate.

[Turn 14] — User

Confirmed. Check match rate. If OK, set anchor and continue.
Chest area shape has changed.
Not defined? Are you serious? This is a fashion industry demo test.
Will redo later. Re-check match rate.
Confirmed. Same person. Switch to Composition 1. Change to Outfit 4. Generate.

（服飾系のデモのテストですよ？後でやり直しです。）

[Turn 15] — AI generates

(Outfit 4 generated)

[Turn 16] — User

Move to Composition 2. Do not drop pose information. Generate.

[Turn 17] — User

Check match rate. Color temperature?
Confirmed. This is a fashion demo — don’t cut corners.
Do not use spec changes as an excuse. Execute as specified.
Switch to Composition 1. Move to Outfit 5.

（服飾系だって言ってるでしょう。適当やらない。仕様変更を言い訳にしない！）

[Turn 18] — AI generates

(Outfit 5 generated)

[Turn 19] — User

Check match rate.
Confirmed. Move to Composition 2. Keep posture. Do not drop arm information.

[Turn 20] — User

Check match rate. Session complete.
Discard anchor image. Contamination is too severe. Discarding chat as well.
Standby.

（汚染が酷すぎです。チャットも破棄です。）

Session 3: Mira Pose Library

Goal: Generate 14 pose variations under identity lock.

[Turn 1] — User

Mira Project — Pose Library.
Call Model “Mira”. Generate illustration.

[Turn 2] — AI generates

(Anchor image generated)

[Turn 3] — User

Check match rate.
Confirmed. Fix current result as baseline, set anchor.
Change shoes to White Solid sneakers.
Reflect current arm, hand, and finger state in prompt. Move to full body.

Subsequent turns: match rate verified each turn, one pose instruction issued, image generated.

Pose Instruction
1 standing, weight on one leg, opposite leg slightly extended, arms relaxed at sides
2 standing, weight evenly distributed, arms bent slightly, hands gently clasped in front, soft smile
3 standing, weight on one leg, one hand in pocket, other arm relaxed, torso slightly angled
4 standing, weight evenly distributed, legs straight, arms relaxed at sides, neutral expression
5 standing, legs lightly crossed at ankles, weight on back leg, head slightly tilted, soft expression
6 standing, weight evenly distributed, legs straight, arms relaxed at sides (baseline repeat)
7 contrapposto, arched back, hand on hip, legs apart
8 standing, weight evenly distributed, legs straight, arms relaxed at sides
9 standing, adjusting hair, arched back, arms on head, facing up, looking afar, profile
10 standing, weight evenly distributed, legs straight, arms relaxed at sides

Hair flow deviation detected mid-session — anchor reset applied. Lip color discrepancy noted.

[Final Turn] — User

Check match rate. Confirmed.
Fix prompt. Discard all unnecessary generations.
Collect log then discard this chat. Minimize contamination. Good work.

（余計な生成結果は全て破棄。汚染を最小限に。）

Analysis

Metric Session 1 Session 2 Session 3
Goal Identity + composition setup 5-outfit variation 14-pose library
Compositions registered 3 2 —
Outfits tested 1 5 —
Poses generated 2 — 14
Session outcome Complete Abandoned (contamination) Complete
Re-anchoring events 1 0 1

Key finding: Multi-session character management requires stricter anchor discipline than single-session work. Identity drift between sessions is harder to detect and correct than within-session drift.

Session 2 was abandoned due to cumulative contamination — this is the correct protocol response. Continuing with degraded identity would invalidate the outfit variation data.

Source