May 19, 2026
Batch 13 (OpenAI) — Film Language A/B Test
- Same Edge B15 prompt as Batch 12 Test C, but with 35mm film + Kodak color science language added
- Testing whether film/camera language suppresses gpt-image-2 texture artifacts
- Compare directly against Batch 12 Test C (extras section) for A/B comparison
- Film language: "Shot on 35mm film, warm analog grain, Kodak color science, macro lens, organic bokeh"
Cost: $0.21 (4 x $0.053)
May 19, 2026
Batch 12 (OpenAI) — NBP Prompt Adaptation
- Adapted 3 proven NBP prompts + their exact reference images for gpt-image-2
- Test A: Down They Swam 2B (4 refs: scene anchor + style + Ollie + Dot)
- Test B: Down They Swam V3 cliff edge (3 refs: style + Ollie + Dot)
- Test C: Edge B15 n_t085 (6 refs: 2 char + 4 scene favorites)
- All via images.edit endpoint, quality=medium, 1536x1024
- Testing whether NBP's best compositions translate to OpenAI model
Cost: $0.64 (12 x $0.053)
May 18, 2026
Batch 11 (OpenAI) — Updated Skill Test
- gpt-image-2-prompt skill rewritten 2026-05-17 after cross-referencing OpenAI official docs
- Fixes: Size param (flexible resolution, not DALL-E 3 fixed), quality default, relaxed positive-only rule, added ref image API code, style transfer, character consistency via refs, n param, streaming, iteration strategy
- Test 1 (V1-V4): Generate endpoint with n=4 in single call, quality=high, 1536x1024, prompt-only (no refs)
- Test 2 (Ref V1): Edit endpoint with 2 reference images: Ollie character ref (illust_ollie_nbp_v3) + style ref (edit08_v3_pick1). Labeled by index and role per updated skill.
- Testing updated skill before reporting results to OpenAI contact
Cost: $0.95 (4 gen + 1 edit at $0.19 each)
May 10, 2026
Batch 10 (OpenAI) — Pure Reef, No Mixed Media
- ZERO mixed media language — no collage, paper mache, gouache, ink, watercolor, diorama, craft
- Pure photorealistic underwater photography prompting, style drawn from batch14/15 character generation prompts
- Only 2 refs: Ollie (batch15 o5_playing) + Dot (batch14 d1_tilted_float) — NO scene ref
- Dot is hidden WITHIN the reef, angled/tilted, active and dynamic — NOT in a pile of shells, NOT facing viewer
- Testing hypothesis: does removing all mixed media language produce better character integration?
Cost: $0.76 (4 x $0.19)
May 10, 2026
Batch 9 (OpenAI) — Refined B8 + Varied Scene Refs
- Addresses Batch 7 failures: restores HIGH quality, 3:2 ratio, scene refs
- Batch 7 root cause: LOW quality ($0.006), 1:1 square, no scene ref
- Two scene refs for variety: b2_v1 (V1-V2) + b2_v3 (V3-V4) from character-and-style batch2
- Same correct char refs as B8: batch15 Ollie (o5_playing) + batch14 Dot (d1_tilted_float)
- Refined prompt: stronger character matching language, explicit anti-texture safeguards
- Same 80/20 photographic/handcrafted ratio, mixed media layering
Cost: $0.76 (4 x $0.19)
May 10, 2026
Batch 8 (OpenAI) — Character&Style Refs + Mixed Media
- Combined B6 mixed-media approach with character-and-style batch2 prompt language
- Scene ref from character-and-style page (b2_v1.png — coral reef mixed media)
- Correct character refs: batch15 Ollie (o5_playing) + batch14 Dot (d1_tilted_float)
- 3D collage coral foreground + photorealistic middle + ink/watercolor background fragments
- "80% photographic reality, 20% handcrafted art" — proven language from character-and-style batch2
- HIGH quality restored (B7 failure was due to LOW quality + no scene ref)
Cost: $0.76 (4 x $0.19)
May 9, 2026
Batches 8-10 (NBP + NB2) — Correct Character References
- Fixed character refs: batch15 Ollie (o5_playing) + batch14 Dot (d1_tilted_float)
- Previous batches 5-7 used wrong character references
- Three approaches retested: refs only (B8), text only (B9), scene+refs (B10)
- 24 images, all free tier
Cost: $0.00
May 8, 2026
Batches 5-7 (NBP + NB2) — Three Reference Approaches
- Batch 5: White-background character refs (no scene bleed)
- Batch 6: Text-only, zero refs (test prompt-only integration)
- Batch 7: Best scene (nb2_b10_v4) as scene ref + white-bg char refs
- Same refined prompt across all three, testing reference strategy only
- 24 images total: 8 per batch (4 NBP + 4 NB2)
Cost: $0.00 (all on free tier)
May 8, 2026
Batch 4 (NBP + NB2) — Fixed Character Descriptions
- Removed contradicting character descriptions that were confusing the model
- Ollie: single clear description matching o2_reaching ref (gold foil iris, stipple texture, ceramic quality)
- Dot: single clear description matching d3_in_motion ref (pearl with internal yellow arch light, NOT robot/screen)
- Same underwater photography blend and scene composition as Batch 3
- NBP: 4 variants (~$0.54), NB2: 4 variants (free)
Cost: $0.54
May 8, 2026
Batch 3 (NBP + NB2) — Collage + Underwater Photography Blend
- Character descriptions pulled from actual generation prompts (gold foil stipple, chromatophore shimmer, pearlescent nacre)
- Added underwater photography reality layer — collage submerged in real ocean water
- Realistic bubbles, caustics on all surfaces, water particle haze
- Updated Dot description: pearlescent white, yellow upside-down half U eyes, internal light
- Same character refs as Batch 2
- NBP: 4 variants (~$0.54), NB2: 4 variants (free)
Cost: $0.54
May 8, 2026
Batch 2 / prev. Batch 8 (NBP + NB2) — Down They Swam Prompt for Hide and Seek
- Prompt adapted from Down They Swam progression page (Page 3 threshold scene)
- Mixed media collage style with specific material definitions
- New character refs: batch4 Ollie (o2_reaching), batch14 Dot (d3_in_motion)
- Updated character descriptions: yellow iris Ollie, tiny sphere robot Dot with green/yellow eyes
- Iridescence restricted to mussel shell interiors only
- NBP: 4 variants ($0.54), NB2: 4 variants (free)
Cost: $0.54
May 8, 2026
Side-by-Side Model Comparison — 3 Prompts x 3 Models
- Direct comparison: same prompt run on
gpt-image-2, NB2, NBP
- Prompt A: NBP Final style (short, artistic direction)
- Prompt B: Detailed material/rendering spec with realism anchors
- Prompt C: Layered media with watercolor washes and translucent forms
- All high quality, prompt only (no reference images)
- 27 images total
Cost: ~$2.91 (OpenAI $1.71, NBP $1.21, NB2 free)
May 8, 2026
Batch 7 (OpenAI) — Low-Cost Iteration, Char Refs Only
- Switched to LOW quality ($0.006/image) for rapid prompt iteration
- Dropped scene reference (was anchoring Batches 5-6 identically)
- Only Ollie + Dot character refs via images.edit
- Mixed-media described as illustration style, not physical materials
- Testing if prompt alone can drive the mixed-media aesthetic
Cost: $0.024 (4 x $0.006)
May 8, 2026
Batch 6 (OpenAI) — Mixed-Media Restored + Polished Surfaces
- Restored mixed-media handcrafted diorama language (Batch 5 was too sterile)
- Materials described as "smooth sculpted" not "torn with fiber texture"
- Brought back coral diversity, color variety, gold foil, dense sea life, decoys
- Same 3 reference images via images.edit endpoint
- Testing hypothesis: mixed-media LOOK without texture-triggering material descriptions
Cost: $0.76 (4 x $0.19)
May 7, 2026
Batch 5 (OpenAI) — Anti-Texture Rewrite + Character References
- Complete prompt rewrite: removed all material texture language
- Used images.edit endpoint with 3 reference images (Ollie, Dot, scene)
- Hypothesis: texture artifacts caused by material descriptions (torn paper, paper mache, grain)
- Test: describe 3D visual aesthetic only, explicit anti-texture instructions
- Batch numbering standardized 1-5, reverse order (newest first)
Cost: $0.76 (4 x $0.19)
May 7, 2026
v3 (All Models) — Texture Mitigation + 35mm Film + Corrected Characters
- Applied gpt-image-2 skill texture mitigation to NBP Final style prompt
- Added 35mm film aesthetic (warm analog grain, Kodak color science, natural depth of field)
- Dot eyes corrected to "two yellow glowing arch lines" (not crescents, not U shapes)
- Ollie char ref updated to 3D NBP v2 character sheet
- Material block restored, SNAIL section added, RENDERING section with full palette
- Generated across all 3 models: NBP (8 images, $1.07), NB2 (8 images, free), OpenAI (4 images, $0.76)
- Tracking notes added to all batches for progress visibility
Cost: $1.83 (NBP $1.07 + OpenAI $0.76)
May 6, 2026
Batch 3 / v2 (OpenAI) — NBP Final Style, Hide and Seek
- Ran v2 prompt (NBP Final style) on gpt-image-2 for Spread 05
- 4 variants at 1536x1024, high quality
- Same prompt used across NBP, NB2, and OpenAI for direct comparison
Cost: ~$0.34 (OpenAI gpt-image-2)
May 6, 2026
v2 Prompt (NB2) — NBP Final Style, Hide and Seek
- Ran v2 prompt (NBP Final style) on NB2 for Spread 05
- NB2: 8 images (4 with D10 refs + 4 prompt-only), free tier
- Same prompt used across NBP and NB2 for direct comparison
Cost: Free (NB2)
May 6, 2026
Batch 1 v2 (NBP) — NBP Final Style, Hide and Seek
- Regenerated Spread 05 with prompt inspired by Section 4 NBP Final (Sunny Reef)
- Used D10 Mixed Media reference images (char ref + style ref)
- Previous v1 images preserved but collapsed
Cost: $1.07 (8 x $0.134)
May 6, 2026
Batch 1 (NBP) — Hide and Seek Spread 05
- Generated 8 images for Spread 05 using Nano Banana Pro
- Two paths: Prompt Only (4 variants) and Scene Reference (4 variants)
- Model:
nano-banana-pro-preview
- Same prompts as NB2 batch for direct model comparison
Cost: $1.07 (8 x $0.134)
May 6, 2026
Batch 2 (OpenAI) — Skill-Optimized Prompts
- Regenerated all 5 spreads with improved prompt structure
- Applied
gpt-image-2 prompt engineering skill
- Material block standardized to first position
- Texture artifact mitigation language added
- All prompts restructured to skill template order
Cost: $3.80 (20 x $0.19)
May 6, 2026
Batch 1 (OpenAI) — 5-Spread Generation
- Generated 20 images across 5 spreads (01, 05, 07, 08, 13)
- Model:
gpt-image-2, quality: high, 1536x1024
- 4 variants per spread
- Spreads 05 and 13 regenerated with fixed character descriptions
- Original pre-fix images preserved for reference
Cost: $3.80 (20 x $0.19)
May 5, 2026
Batch 1 (NB2) — Initial Test
- Generated 16 images across 2 spreads (05, 13)
- Model:
Nano Banana 2 (gemini-2.0-flash-exp), 1024x512
- Two paths tested: Prompt Only (text prompt) and Scene Reference (with ref images)
- 4 variants per path per spread