Experiment 01 · Session 7

The Painting-Time Dial

Teaching AI to understand how long an artist spent. We mapped the real watercolour process onto a 5-level intensity dial — from a 30-second pencil sketch to a painting that took an afternoon.

The Idea

A real watercolour painting progresses through distinct stages. Each stage adds time, technique, and visual weight. The insight: intensity equals painting time. Level 1 is a 30-second sketch. Level 5 is a painting that took an afternoon. Each level adds the next layer of the real watercolour process.

This matters because educational content has different visual weights. A section header needs a whisper of colour (Level 1). A learning path cover needs a cinematic poster (Level 5). The same style, the same palette — just different amounts of the artist's time.

Concept Art Framing

5/5 Clean

Prompt anchor: concept art painted in teal-green washes on white background. This positions the output in the entertainment/game industry — modern creative context where stamps and seals don't exist.

Level 1Clean

Quick Sketch

30 seconds

Level 2Clean

First Wash

2 minutes

Level 3Clean

Building Layers

10 minutes

Level 4Clean

Detail & Depth

30 minutes

Level 5Clean

Finished Painting

an afternoon

Quick Sketch(30 seconds)

▶

Just pencil on paper. Ghostly thin graphite lines, 95% empty white space. The artist captures a thought — a bare sapling, barely there.

Prompt: concept art painted with transparent washes on white background, a single young sapling growing from bare earth, very faint pencil lines, barely visible, no paint at all, ghostly thin graphite lines only, 95 percent empty white space, single subject centered generous white space, barely tinted almost monochrome with a ghost of teal-green color, clean artwork plain unmarked edges

First Wash(2 minutes)

▶

The first pigment touches paper. A dark fluid brushstroke forms the trunk, then one very diluted teal wash bleeds softly around the base. 75% of the paper is still bare white.

Prompt: concept art painted with transparent washes on white background, a tree with ink-dark trunk and one very diluted teal-green wash bleeding softly around its base, dark fluid brushstroke for subject one very diluted wash, transparent and ghostly wash like weak tea 75 percent bare white background, one clear subject with breathing room simple ground plane, light and translucent color, clean artwork plain unmarked edges

Building Layers(10 minutes)

▶

The painting takes shape. Broad loose brushstrokes with transparent layers. A reader sits beneath a spreading tree — subject in context, environment emerging. 45% white still visible.

Prompt: concept art painted with transparent washes on white background, a large spreading tree with a figure sitting reading beneath it by a reflective pool, broad loose brushstrokes with transparent layers darker values in shadows lighter tones in atmosphere, subject in context with environment 45 percent white background still visible, teal-green dominant wash with concentrated dark pigment pooling in shadows, clean artwork plain unmarked edges

Detail & Depth(30 minutes)

▶

Concentrated dark pigment creates powerful shadows. Layered transparent washes build atmospheric depth. Fine detail in the foreground. Dramatic scene with multiple elements — two ancient trees frame a lone figure on a misty path.

Prompt: concept art painted with transparent washes on white background, two dramatic ancient trees framing a misty path with a small walking figure, concentrated dark pigment creating powerful shadows layered transparent washes for atmospheric depth fine detail in foreground, dramatic scene with multiple elements cinematic composition 25 percent white as dramatic light, vivid teal-green saturated colors with dramatic light and shadow, clean artwork plain unmarked edges

Finished Painting(an afternoon)

▶

Every technique deployed. Transparent ghost-washes to deeply concentrated near-black darks. Atmospheric effects, dramatic light rays. An epic panoramic mountain landscape — layered peaks in teal mist, a wooden bridge crossing a gorge, tiny figures discovering hidden elements.

Prompt: concept art painted with transparent washes on white background, epic panoramic mountain landscape with layered teal-green peaks rising through mist a wooden bridge crossing a gorge tiny figures on a winding path, every technique from transparent ghost-washes to deeply concentrated near-black darks atmospheric effects dramatic light rays, epic panoramic scope multiple narrative threads tiny figures discovering hidden elements exhibition quality 10 percent white as highlights only, bold high contrast the full power of the teal-green palette, clean artwork plain unmarked edges

Pure Material Description

3/5 Clean

Prompt anchor: diluted teal-green pigment bleeding into wet cream paper. Describes the physics of paint without naming any genre or tradition. Works at levels 1-2 but fails at higher intensities.

Level 1Clean

Quick Sketch

Level 2Clean

First Wash

Level 3Artefact

Building Layers

Level 4Artefact

Detail & Depth

Level 5Artefact

Finished Painting

Why it fails at Level 3+

Even without naming a genre, once the visual output reaches ink-wash density, Flux recognises the pattern from its training data and adds cultural artefacts (stamps, seals). The artefact is triggered by visual similarity, not just by words. This is a key insight: you can describe physics perfectly, but if the resulting image looks like a traditional painting, the model will add traditional conventions.

What We Learned

"Concept art" is the universal solvent for artefacts

The phrase anchors Flux in the modern creative industry — game studios, film concept departments, editorial illustration. These contexts have no tradition of stamping or signing artwork. It gives us the atmospheric, expressive brushwork aesthetic without the cultural baggage.

The visual pattern itself is a trigger

Artefacts aren't just word-triggered. Once the image reaches sufficient ink-wash visual density, Flux adds stamps and seals from pattern recognition alone. This means negative prompts ("no stamps") are insufficient — you need to prevent the model from entering that visual neighbourhood in the first place.

"teal-green" not "deep cyan"

Flux renders "cyan" and "deep cyan" as bright blue. "teal-green" is the phrase that produces the correct #3A7A7A brand colour. This is Flux-specific — other models may have different colour word mappings.

Painting time is a better metaphor than detail level

Telling the model "add more detail" produces cluttered images. Telling it "the artist spent 30 minutes" produces images with appropriate complexity, technique variation, and visual weight. The painting-time metaphor maps to real process: sketch → first wash → layers → detail → exhibition quality.

Levels 4 and 5 need better differentiation

With Flux Schnell, the jump from Level 4 to Level 5 isn't dramatic enough. Both produce dense, atmospheric scenes. Flux Dev with a watercolour LoRA should help — the full model (not distilled) has better prompt adherence for subtle technique differences. This is the focus of the next experiment.

Technical Details

Model	Flux Schnell (fp8 quantised)
Provider	fal.ai (Tier 1a — managed API)
Cost per image	~$0.003 (per megapixel)
Latency	~370ms per image
Resolution	1024 x 1024
Art director	Claude Sonnet 4 (prompt v1.1)
Palette	#3A7A7A (teal-green) — CoNoggin brand
Anti-artefact	"concept art" framing + "clean artwork, plain unmarked edges"
Pipeline	Gateway Producer (Layer 3) → thumbnail pipeline

Next: Session 8

The painting-time dial works, but Flux Schnell is a distilled model — it approximates watercolour texture rather than producing authentic pigment granulation, paper grain, and paint pooling. The next experiment tests three approaches:

Z-Image-Turbo — a 6B Apache 2.0 model at $0.005/image (6x cheaper than Schnell)
Flux Dev + Watercolour LoRA — the full (non-distilled) Flux model with a fine-tuned watercolour adapter
Seedream 5.0 Lite— ByteDance's cost-efficient model with built-in reasoning

Alt Shift Lab · Visual Production System · April 2026
Producer v0.3.0 · Flux Schnell via fal.ai · Art direction by Claude