The Model Shootout
Four AI image models run the same five painting-time prompts. Same subject, same teal palette, same concept-art framing — different architectures, different costs, different results. Which model produces the best watercolour illustrations for educational content?
Test Design
Each model receives the exact same 5 prompts from Session 7's painting-time dial — the "concept art" framing that eliminated artefacts. The prompts describe a tree/landscape scene at escalating intensities from a 30-second pencil sketch to an afternoon masterwork. All use the CoNoggin teal palette (#3A7A7A).
Side-by-Side Comparison
Quick Sketch
30 seconds
Beautiful minimal sapling with roots. Great whitespace.
Coloured sprout — too much paint for a "30-second sketch". More illustration than pencil.
Tiny delicate sapling. Perfect whitespace, genuinely looks like pencil on paper.
Ghost-faint sapling. Very minimal — perhaps too blurry. LoRA softens too much at low detail.
First Wash
2 minutes
Single dark trunk, diluted teal wash. Clean.
Splattery digital watercolour tree. Decent but feels processed.
Dark ink trunk with minimal wash. Best Level 2 of any model — "first wash" perfectly captured.
Soft teal wash tree. The bleeding edges are genuinely watercolour, but the teal has shifted cyan.
Building Layers
10 minutes
Reader under spreading tree — classic Schnell quality.
Tree with reader and reflection. Good but lacks paint texture.
Reader under tree by pool. Gorgeous wash bleeding. The most authentic watercolour.
Tree with reader. Authentic wash quality! But over-softened and colour drifting toward turquoise.
Detail & Depth
30 minutes
Dramatic gnarled trees, misty path. Strong composition.
Moody forest path — the best Z-Image level. Paint bleeds over frame edge.
Two ancient trees framing a path. Stunning bark detail, atmospheric mist.
Two trees with figure — dreamy but TOO soft. Small signature mark bottom-right. Cyan shift.
Finished Painting
an afternoon
Epic panorama BUT "© 12014" text in bottom-right corner.
Mountain panorama. Smaller scale than Schnell, edges painted over white.
Panoramic mountains with bridge and light rays. Best Level 5 — real depth, no artefacts.
Mountain bridge panorama. Best watercolour texture of the set — real wash layers — but blurry.
Model Scorecards
Flux Schnell
BalancedThe reigning baseline. Fast, cheap, good prompt following. But the fresh run revealed a text artefact ("© 12014") at Level 5 — showing the concept art framing isn't 100% reliable. Watercolour texture is approximated, not authentic.
Z-Image-Turbo
Speed WinnerThe speed and cost champion. 6x cheaper than Schnell with comparable latency. But the watercolour texture reads as "digital illustration with watercolour filter" rather than authentic paint. Edges bleed over the frame at high intensities. Strong for quick previews.
Seedream 5.0 Lite
Quality WinnerThe quality champion. The most authentic watercolour feel of any model tested — genuine pigment pooling, paper texture, expressive brushwork. Best intensity differentiation (Level 2's stark ink trunk is stunning). Clean across all 5 levels. But 40-50x slower than Schnell and 12x more expensive, with no LoRA customisation path.
Flux Dev + LoRA
Best TextureThe LoRA delivers genuine watercolour texture — wet-on-wet bleeding edges, wash softness that Schnell can't produce. But the images are over-softened (LoRA strength 0.85 is too high). The teal shifts toward bright cyan/turquoise. Level 4 has a small signature artefact. The path forward: reduce LoRA strength to 0.5-0.7 and adjust colour prompting.
Cost at Scale
| Model | Per Image | 100 images | 1,000 images | Time for 100 |
|---|---|---|---|---|
| Z-Image-Turbo | $0.005 | $0.50 | $5.00 | ~2 min |
| Flux Schnell | $0.003 | $0.30 | $3.00 | ~1.5 min |
| Seedream 5.0 Lite | $0.035 | $3.50 | $35.00 | ~75 min |
| Flux Dev + LoRA | $0.033 | $3.30 | $33.00 | ~57 min |
Key Findings
Seedream produces the most authentic watercolour — but you can't customise it
Seedream 5.0 Lite's Level 2 (the stark ink trunk with a single diluted wash) is the most convincing watercolour output of any model tested. The paint texture, pigment pooling, and paper grain are genuinely painterly. But there's no LoRA support, no local weights, and a proprietary license. You get what ByteDance gives you.
Z-Image-Turbo is a revelation for previews and bulk generation
At $0.005/image and 1.3s latency, Z-Image-Turbo is absurdly efficient — 6x cheaper than even Schnell. Apache 2.0 license, LoRA support on fal.ai, ComfyUI checkpoint available. The watercolour quality is the weakest of the four, but for "quick preview before the user picks which one to regenerate at production quality" it's perfect.
The watercolour LoRA works but needs tuning
The gokaygokay/Flux-Watercolor-Strokes-LoRA at strength 0.85 produces genuine wet-on-wet bleeding edges and wash softness — things Schnell literally cannot do. But it over-softens the image and shifts the teal toward cyan. Next step: reduce to 0.5-0.7 strength and add explicit colour correction in the prompt. The LoRA's trigger words (WTRCLR watercolor) may also be competing with our concept art framing.
Flux Schnell's concept art framing isn't bulletproof
The fresh Level 5 generation produced a "© 12014" text artefact in the corner — something the Session 7 run didn't show. This means the anti-artefact strategy is probabilistic, not deterministic. At production scale (thousands of thumbnails), a small percentage will have text contamination. A post-generation quality check or a text-detection filter may be needed.
No single model wins on all axes
Speed: Z-Image > Schnell >> Seedream ≈ Flux Dev. Quality: Seedream >> Flux Dev > Schnell > Z-Image. Cost: Schnell > Z-Image >> Flux Dev ≈ Seedream. The right answer is a tiered approach — fast model for previews, quality model for production assets.
Recommended Strategy
The tiered approach: add a quality parameter to the thumbnail pipeline.
- quality: "fast" — Flux Schnell ($0.003, ~0.8s). For bulk generation, course import, auto-thumbnails during content creation. The current production default.
- quality: "preview" — Z-Image-Turbo ($0.005, ~1.3s). For generating 3-5 options for user selection. Apache 2.0, LoRA-ready if we want to fine-tune later.
- quality: "production" — Seedream 5.0 Lite ($0.035, ~45s). For final hero assets — learning path covers, marketing materials, brand-defining images. The quality speaks for itself.
The LoRA path remains open. Flux Dev + watercolour LoRA at a lower strength (0.5-0.7) could become the production tier once tuned. It offers customisation that Seedream can't — org-specific style LoRAs, IP-Adapter for reference images, ComfyUI workflow flexibility. This is a Session 9 tuning task, not blocked.
Technical Details
| Date | April 16, 2026 |
| Prompt set | Session 7 painting-time dial (concept art framing, 5 levels) |
| Palette | #3A7A7A (teal-green) — CoNoggin brand |
| Resolution | 1024 x 1024 (all models) |
| Flux Schnell | fal.ai direct API, fp8 quantised |
| Z-Image-Turbo | fal.ai direct API, 6B DiT, Apache 2.0 |
| Seedream 5.0 Lite | fal.ai (bytedance/seedream/v5/lite/text-to-image) |
| Flux Dev + LoRA | Replicate comfyui/any-comfyui-workflow (L40S), flux1-dev-fp8 + gokaygokay/Flux-Watercolor-Strokes-LoRA @ 0.85 strength, 25 steps, cfg 3.5, euler/normal |
| LoRA trigger | WTRCLR watercolor (prepended to prompt) |
| Total cost | ~$0.40 for all 20 images |
| Total time | ~12 minutes (including cold starts) |
Alt Shift Lab · Visual Production System · April 2026
Gateway v0.3.0 · Producer Layer 3 · Art direction by Claude