DeepInfra image test

★ Product photoshoot: Kontext vs the current RunningHub

One real seller photo (Hugo cologne, shot in a shop) → professional shots. Both engines run live on the same photo. The hard part: keep the exact bottle and the "HUGO" label.

INPUT · seller's phone snap

Hugo "Man" bottle held in a cluttered perfume shop, mixed lighting, busy background. This is what a Blinco seller actually uploads.

DeepInfra · FLUX.1-Kontext $0.0112/shot · 6–16s · you describe the scene

✓ Bottle + red HUGO label preserved. ✓ Any scene you can describe (prompt-controlled). ✓ Same API as the site imagery, no coins/3rd-party queue.

RunningHub · current 9 shots · 58s · 11 coins + $0.07 cash (measured) · zero prompt

✓ Turnkey — one graph, no prompting, a curated 9-shot spread. ✓ Best label fidelity (HUGO BOSS legible). ✗ Fixed aesthetic, no scene control, depends on RH service + coins.

Verdict — cost is basically a tie, and control breaks the tie. With RHcoin valued at $0.00036 (derived from RunningHub's own $26 plans), measured on the same photo:
• RunningHub = 11 coins ($0.004) + $0.07 wallet = $0.074 for a fixed 9 shots (~$0.008/shot)
• Kontext = $0.099 for 9 shots (~$0.011/shot) — but you choose the count: 4 shots @768px ≈ $0.03
Per 1,000 sellers: RunningHub's fixed-9 = $74; Kontext at the 4 shots a seller actually needs = $32. So Kontext is cheaper in practice because it doesn't force 9. Quality is a wash (RunningHub's label fidelity is a hair better; Kontext's scenes are cleaner e-commerce). And RunningHub's dual currency is a trap: premium runs drain the USD wallet, not the coins — this account has 72k coins but only ~$7.8 wallet, so it's ~111 runs from a forced top-up with 70k coins stranded. Kontext = one key, one balance, prompt-controlled scenes, fewer shots.

My take: move product shots to Kontext for control + one-vendor simplicity, and keep RunningHub as a fallback. Want me to wire Kontext into productShots.ts (same async upload→edit→store shape it has now, with 3–4 scene prompts per product)?

Can product shots be cheaper than Kontext?

Short version: Kontext-dev (~$0.01/img ≈ $10 per 1,000) is already the cheap floor for real fidelity. Going cheaper breaks the product.

✗ $0.0005 · schnell img2img 20× cheaper

Wrong color (green), wrong shape, label now reads "HUGE". It invents a generic bottle, not the seller's. Unusable.

✓ ~$0.011 · Kontext-dev the sweet spot

Exact bottle, frosted blue, silver cap, red HUGO label preserved. Real product, clean scene.

Image-to-image model	Per image	Per 1,000	Keeps the real product?
schnell img2img	$0.0005	$0.50	✗ No — reinvents it ("HUGE")
FLUX.1-Kontext-dev	~$0.011	~$10	✓ Yes — best value
FLUX-1-Redux-dev	$0.012	$12	~ variation, not instruction-edit
Qwen-Image-Edit	$0.028	$28	✓ Excellent, a touch cleaner
Qwen-Image-Edit-Max	$0.075	$75	✓ (overkill)

Answer: $40/1,000 was a high estimate — Kontext-dev is ~$10/1,000, and it's already the cheapest model that keeps the seller's actual product. The only thing cheaper (schnell img2img, $0.50/1,000) hands back a different bottle with a garbled label — worthless for a real store.

The right way to cut cost isn't a cheaper model, it's fewer/smaller shots: a seller needs ~3–4 hero shots once, not 9 — and at 768px with fewer steps each drops to ~$0.006–0.008. That's ~$0.02–0.03 per seller onboarded. Cost scales per seller, not per 1,000 images. So the whole product-shoot bill for a seller is a rounding error either way.

FLUX-1-schnell — the cheap one

10 realistic briefs · $0.000375 each · $0.00375 total for all ten

schnell vs dev — is 18× the price worth it?

Same prompt, same seed-free run. schnell $0.000375 vs dev $0.00675. Judge for yourself.

SaaS / product UI — the honest answer

Can it do dashboards & app screens for mockups? Partly. The rule: models can render the shape of a UI but not legible text. So it depends entirely on how the UI is shown.

✓ WORKS Device mockups & product shots — UI is small/incidental, gibberish invisible

⚠ WORKS w/ care Abstract 3D / tech heroes — great, but add "no text, no letters" or it hallucinates garbled words (see the "SasaS")

✗ FAILS Full-bleed UI screenshots — structure is right, but every label & number is illegible scribble. Unusable as foreground.

So, for Blinco's SaaS sites: use FLUX for the device mockups and 3D heroes (phone-in-hand, laptop-on-desk, glass-card abstracts) — those are excellent and cheap. But do NOT use it to fake a crisp app screenshot: the text turns to mush. For a believable dashboard/app screen the right move is to render a real mini-UI in HTML/CSS inside a device frame (Blinco already builds HTML — a "product frame" block draws perfect, legible UI deterministically). That beats any AI screenshot and costs nothing. Same limitation on FLUX-dev, Midjourney, DALL·E — it's the whole category, not the cheap tier.

The "one canvas → crop" trick, tested on all 3 ideas

Smart instinct. It's a trap for photos (per-megapixel pricing means no saving) and a gem for assets (small + reusable + transparent).

✗ Photos / e-commerce Kontext can't tile 4 scenes; and a 4× canvas costs 4× — the pixels cost the same however you arrange them.

One 1440² canvas → 4 crops @720px ($0.001). Problems: all four inherit one global look (everything went dark — no bright bakery next to a dark gym), each crop is only 720px (soft for a hero), framing gets clipped, and it's the same price as 4 separate images. You lose control + resolution for zero saving.

Verdict: for section heroes, generate each separately at slot size — it's already ~$0.0004 and you get independent art-direction. The grid only helps for a row of matching thumbnails.

✓ Floating assets / 3D / vectors This is the real win — one sheet → many reusable, transparent assets the AI drops into sections.

One 1024² sheet = 9 assets for $0.0005 ($0.00006 each). Crop with code (free) → remove background → drop into any section. Works because assets are shown small (a 341px tile is plenty) and are reused across every site — build the library once.

→

→

Transparency: Bria/remove_background is robust but $0.018/asset (dominates cost). Cheaper: generate the sheet on a pure chroma background and key it out with code (sharp) — free. 16 transparent assets for ~$0.0005 total. DeepInfra also has Bria-3.2-vector for true SVG-style vectors (infinite scale, tiny files) — worth a follow-up.

Net: Don't crop photos — per-megapixel pricing means the grid saves nothing and costs you control + resolution; generate heroes individually (fewer, at slot size). Do build an asset library the sheet way: one schnell sheet of 9–16 floating 3D/UI/vector assets on a chroma background → crop → key out → a reusable, transparent, on-brand asset set for ~$0.0005, that the Forge AI can scatter into sections as floating ornaments. That's the genuinely clever half of the idea.

The numbers

Model	Cost / image	Avg speed	Per Blinco site (~8 imgs)	Per 1,000 sites
FLUX-1-schnell	$0.000375	~1.0s	$0.0030	$3.00
FLUX-1-dev	$0.006750	~2.4s	$0.0540	$54.00

Recommendation. Go with FLUX-1-schnell as Blinco's image engine. Across 10 varied briefs it never failed and never produced garbage — the food and interior shots are indistinguishable from stock photography, and it interprets art-direction cues (golden hour, rim lighting, shallow depth of field) correctly. dev costs 18× more and reads as more "artistic," but it isn't more usable for business heroes — on the gym brief it went full silhouette while schnell gave a clean, on-message hero. At $0.003 per site this is effectively free, and unlike the current Pexels stock it's unique per business and unlike Vertex it has no quota wall.

Next step (your call): wire schnell into stages/imagery.ts as the primary provider, ahead of Pexels/Vertex. I have not touched the live engine yet — this was a read-only test.