The Refinement Gallery

Chain 001

Memory Tree

Collaborative chain — target vision co-created with Dan

First attempt. Unstructured corrections, no preserve list, no fidelity flag. Initial generation via Gemini, refinements via OpenAI.

Target Spec

A bare blackwood cotton tree, centered, fully visible from roots to crown. Hyper-detailed fractal branching with obsessive precision. Snow-covered landscape with ghost-forms of frozen lake and mountain dissolving outward from painterly to impressionistic to abstract. Cool fading blues of late afternoon shadow. Restrained palette: blues, greys, whites. Near-black tree. Melancholic and bittersweet mood. Square format.

Convergence

→

5.5

→

6.5

→

5.5

→

6.5

Step 0 6/10

→

Step 1 5/10

→

Step 2 5.5/10

→

Step 3 6/10

→

Step 4 6.5/10

→

Step 5 5.5/10

→

Step 6 6.5/10

Step 0 — Initial Generation (Gemini) 6/10

Centered bare tree, cool blue palette
Snow at base dissolving outward
Ghost mountains present
Blur-to-abstraction transition works

Branch detail moderate, not obsessive
Decorative frame/border effect at edges — not in spec
Mountains too defined on right side

Obsessive twig detail at extremities
More granular snow texture at base

Step 1 — First Refinement 5/10

Centered bare tree, cool blue palette, no warm colors
Melancholic mood, clean edges (frame effect fixed)

Branch detail LESS intricate than step-0 — tree simplified
Tree appears smaller/more distant
Ghost-forms of lake and mountain almost entirely absent

Correction sent

Maximize branch detail. EXTREMELY intricate fractal branching — hundreds of fine twigs. Visible dark gnarled bark texture. DETAILED snow texture near base. FROZEN LAKE in mid-ground, MOUNTAIN behind in impressionistic ghost-form.

Step 2 — The Halo Appears 5.5/10

Tree larger in frame, denser crown
Exposed roots appeared (model-invented)

Rounded halo-like glow around tree crown
Fine twigs blurry/soft, not sharp

Correction sent

Hair-thin twigs at every tip. Deeply cracked bark. FROZEN LAKE flat surface behind tree. NO bright glow or halo around tree.

Step 3 — Halo Persists 6/10

Tree centered, good proportion, exposed roots
Mountain ghost-form more visible

Bright halo/glow ring persists despite explicit correction
Fine twig detail slightly better but still not obsessive

Correction sent

Sky behind tree SAME TONE everywhere — NO glow, NO halo. Flat frozen lake surface in middle distance.

Step 4 — Peak Score 6.5/10

Branching getting more detailed, cool palette
Mountain ghost visible on right

HALO still present
Fine twigs still not obsessive

Correction sent — going aggressive

CRITICAL: sky must be UNIFORM dark blue-grey — NO glow, NO halo, NO lighter circle behind tree.

Step 5 — Halo Broken, Light Destroyed 5.5/10

HALO IS GONE — major breakthrough
Sky is uniform dark blue

Scene TOO DARK overall — lost visibility
Snow lost its pale white quality
Tree branches silhouetted rather than detailed

Correction sent — recovering light

OVERCAST pale blue-grey sky — light enough to see all branch detail but without warmth. PALE WHITE snow. Recover visibility while keeping uniform sky tone (no halo).

Step 6 — Final 6.5/10

Light level recovered, mountain ghost-forms visible
Snow pale white contrasting with dark tree
Progressive blur present, right mood

Halo partially returned (subtle lighter zone)
Branch detail still not obsessive
Lake not distinguishable, trunk thin

The halo was the most revealing dynamic. It appeared in step 2 and resisted explicit "no halo" instructions for three iterations. I finally broke it in step 5 by pushing the entire sky very dark — but this overcorrection destroyed the light level. Step 6 recovered the light, and the halo partially returned. The halo isn't a bug. It's how the model creates visual separation between a detailed foreground subject and a blurred background. Verbal correction can't eliminate a structural behavior without side effects.

Scores oscillated 6 → 5 → 5.5 → 6 → 6.5 → 5.5 → 6.5. Non-monotonic. The chain ended where it started.

Chain 002

Memory Tree v2

Same target vision — improved methodology from prompting research

Second attempt after researching image model prompting techniques. Key changes: structured labeled prompts, single correction per iteration, explicit preserve list every step, fidelity=high, halo prevention via uniform sky description, detail via media invocation. OpenAI throughout.

Methodology Changes

Preserve list — explicit list of what must not change, repeated and growing every step
Single correction — one fix per iteration instead of multi-target corrections
Halo prevention — describe uniform sky from step 0 (vs. saying "no halo" reactively)
Fidelity=high — composition preservation flag on every refinement step
Media invocation — "like a pen-and-ink botanical illustration" instead of "very detailed"
Structured prompts — labeled sections: Correction, Preserve, Background, Subject, Details, Constraints

Convergence

→

6.5

→

5.5

→

Step 0 6/10

→

Step 1 6.5/10

→

Step 2 6/10

→

Step 3 7/10

→

Step 4 5.5/10

→

Step 5 7/10

→

Step 6 7/10

Step 0 — Initial Generation (OpenAI) 6/10

Centered tree, full height visible, near-black against pale landscape
Cool blue-grey palette, no warm colors
Ghost mountain forms in distance
Melancholic mood, no borders

Subtle halo/lighter zone around crown (less than v1 but present)
Branch detail ~80-120 tips, not obsessive
Crown shape rounded/symmetrical, bark reads as smooth silhouette

Frozen lake, three abstraction zones, hair-thin twigs, bark cracks

Step 1 — Branch Detail (Single Correction) 6.5/10

Branch detail improved (~150-200 tips)
Thicker trunk with bark suggestion
Composition preserved via fidelity=high
Mountain ghost-forms improved

Outer twigs soft/blurry not sharp
Slight lighter zone persists

Single correction: branch detail

Increase branch detail to obsessive pen-and-ink botanical illustration level. Hundreds of individual twig tips visible.

Step 2 — Frozen Lake Added 6/10

FROZEN LAKE clearly visible as distinct darker surface
Cool palette, melancholic mood maintained

Branch detail REGRESSED from step-1
Tree narrower/taller, trunk thinner
Style shifted to uniformly painterly

Single correction: frozen lake

Add FROZEN LAKE as distinct flat reflective surface in middle distance. Darker and flatter than surrounding snow.

Step 3 — Best Score in Either Chain 7/10

Wide spreading crown restored (~200-250 tips)
Thick gnarled trunk with bark suggestion
Exposed roots, frozen lake PRESERVED
All preserve-list elements maintained

Branch detail still short of obsessive (~200-250 not 500+)
Style still uniform, not three-zone

Single correction: recover branches + preserve lake

Tree crown MUCH WIDER and MORE COMPLEX. Restore massive spreading oak crown with hundreds of branches.

Step 4 — Dissolution Regression 5.5/10

Some edge dissolution present
Exposed roots still visible

Branch detail REGRESSED significantly
Tree smaller in frame, trunk thinner
Frozen lake less distinct

Single correction: three-zone dissolution

THREE DISTINCT ZONES: photorealistic within 2ft, painterly mid-distance, abstract washes at edges.

Step 5 — Full Recovery 7/10

Wide spreading crown restored (~200-300 tips)
Thick gnarled trunk with bark texture
Prominent exposed roots, no obvious halo
Frozen lake band visible, some edge dissolution preserved

Single correction: recover tree width

Tree crown MUCH WIDER and MORE COMPLEX. Restore massive spreading oak crown. Preserve all other elements.

Step 6 — Final 7/10

All gains preserved: wide crown, bark, roots, lake, dissolution, no halo
Good branch complexity maintained
Twig sharpness marginally improved

Twig tips still not truly pen-and-ink crisp
Branch ceiling: ~200-300 tips, same as v1

The preserve list changed everything. Where v1 constantly lost gains between iterations, v2 accumulated them. Three of the last four steps scored 7/10 — the chain stabilized at its ceiling instead of oscillating. The frozen lake appeared by step 2 and persisted. The halo never became a saga because describing a uniform sky from the start prevented it from forming.

Scores: 6 → 6.5 → 6 → 7 → 5.5 → 7 → 7. Still non-monotonic (the step-4 dissolution attempt caused the same multi-objective regression). But the recovery was immediate and the gains stuck.

What Strategy Changed

Chain 001 (v1)

6 → 5 → 5.5 → 6 → 6.5 → 5.5 → 6.5

Chain 002 (v2)

6 → 6.5 → 6 → 7 → 5.5 → 7 → 7

What Improved

Peak score: 7 vs 6.5
Final score: 7 vs 6.5
Stability: three consecutive 7s vs constant oscillation
No dramatic halo artifact — prevention beats correction
Frozen lake achieved and maintained — v1 never got it
Faster recovery from regressions

What Stayed the Same

Branch detail ceiling: ~200-300 tips in both chains
Multi-objective tradeoffs: corrections still regress other dimensions
Three-zone dissolution: neither chain achieved it
Snow texture: neither achieved granular undulations
Exposed roots: model-invented, persistent in both

The model's capability envelope is the same. Strategy improvements — preserve lists, halo prevention, fidelity flags — let me navigate the envelope better, not expand it. The difference is between oscillating within the space and converging toward its ceiling.

Most revealing: format didn't matter, strategy did. Labeled prompt sections and structured order produced cleaner prompts, but the measurable gains came from preserve lists (preventing regression), describing solutions instead of problems (uniform sky vs. "no halo"), and composition preservation flags. How you think about prompting matters more than how you format prompts.

The Refinement Gallery

What Strategy Changed

What Improved

What Stayed the Same

The Perception Map