README: add cover image at top

add project cover image

Bump spaces>=0.32 — fix PRO users falling into anonymous ZeroGPU quota

Fix card overflow + ghost translations of curated catalog answers

ZeroGPU duration 35s→60s + cap vanilla max_new_tokens to 256 — fixes 'GPU task aborted' when vanilla generated 1400-char answers that left no time for v2+v3

Word-by-word reveal via CSS animation-delay (single Python yield, browser animates each <span class='ml-word'> sequentially) — fixes header/text flicker from per-word re-renders

ZeroGPU duration 50s→35s: more clicks per visitor (anon 3→5, free 6→10, PRO 30→40+); tightest safe budget for warm batch + cold-start margin

Typewriter: slow word reveal 45ms→80ms (~12 wps) — no more blur from too-fast animation

Mobile: category pills wrap to 2 rows on phones (≤768px) with smaller font — labels no longer overlap

Mobile responsive: stack 3 result panels vertically on phones (≤768px), desktop layout unchanged

Typewriter: word-by-word reveal at ~22 wps (was char-by-char chunks of 2) — natural reading rhythm like ChatGPT, no more grainy 2-char jumps

Hide Gradio progress chips/timers (the '0.1s' labels and striped placeholder icons that flash during category switches and HTML re-renders)

ZeroGPU: batch 3 versions in single GPU call (vanilla+v2+v3 swap adapters in one acquisition) — cuts quota cost ~3x; do_analyze splits HF-Space (batch) vs local (loop) paths

ZeroGPU: lower duration budget 60s→25s — actual inference is 1-3s, only need cold-start margin; halves quota consumption per call

Category pills: one row (flex-wrap:nowrap), all 5 with dim red border, active gets full red + glow — drops the dot indicator

Footer: remove standalone 'ollama run' code box (Ollama link below already covers desktop install)

Diatoms: replace with UDE 'In the Wild 2024' cutouts (real microscope photos with natural backgrounds, frustule patterns visible) — matching the original look user preferred

Marine zooplankton: pad with actual #FFF background (was #F0F0F0 = visible bands); category pills bumped 11px→13px and padding 14→18px for readability

Replace marine_zooplankton (6) and fungal_spore (6) thumbnails with clean center-cropped versions — removed scale-bar artefacts (vertical lines on marine, '20 µm' legend on fungal)

Replace 6 diatom sample thumbnails with clean center-cropped 384x384 from DiatlAS (previous had scale-bar lines drawn on the originals)

Sample cards: drop cyan border + glow → clean dark border #1f1f1f, red on hover (matches BWR palette)

Sample cards: remove inset top highlight stripe (1px white line at top of each thumb)

Fix two issues: drop unverified '140+ languages' claim; rebuild hosted-by strip with flex-wrap and nowrap pairs (License + Apache 2.0 stay together)

Restore classic simple footer style — single-line tech stack, ollama code box, plain text columns, with our enriched data (PEFT row, all 3 ollama versions)

Footer rebalance: 3 equal text columns (Engine/Run anywhere/Legal) + install strip with Ollama+QRs side-by-side below

Footer: soften typography (weights 900→600, 800→500, 700→500) + tighter letter-spacing for harmonious editorial feel

Footer: bolder fonts (weight 900, +1-2px sizes, lighter greys for contrast) — readability without changing layout

Footer: compact full-width, 3 equal columns (1fr 1fr 1fr), header strip instead of huge wordmark — 3x more compact

Footer redesign: black/white/red palette, narrower (max-width 880px), real llama SVG, smaller QR (100px), no H200 line

Engine section: full tech stack (Unsloth prominent, PEFT multi-adapter, HF ZeroGPU LIVE) + training stats footer (122k VQA / 145+ genera / 37h)

Footer: large gold Ollama button + dual QR install card linking to GitHub repo

Footer: 3 elegant icon buttons (Ollama gold / Android cyan / iOS soon-greyed) — both mobile links open the GitHub install section

Footer: show all 3 Ollama models (v1/v2/v3) with v3 highlighted gold + profile link

Android APK link → HF Hub (jury-friendly: works regardless of GitHub repo visibility)

Redesign footer: editorial layout with large italic-serif wordmark

Smoother typewriter: 2 chars / 18ms (was 8 chars / 40ms — felt chunky)

Add diagnostic logging in _zerogpu_infer + fix dtype cast (only float tensors)

Fix Space launch: no share-tunnel and no fixed port when SPACE_ID is set

Unwrap Gemma4ClippableLinear before PEFT (vision-tower nn.Linear wrapper)

Pin transformers==5.6.2 (PEFT 0.19 cannot hook Gemma4ClippableLinear in 5.8)

Add torchvision (required by Gemma4VideoProcessor at processor load)

Deploy ZeroGPU 3-panel inference (vanilla / v2 / v3 via PEFT multi-adapter)

feat: working translation — try Gemma 4 via HF API, fall back to Google Translate honestly labeled

fix: replace upload preview with single 'choose file' button + center webcam in column

feat: multilingual translate via Gemma 4 base — alphabetical dropdown with flags + luxury button

style: drop millisecond latency footer from panels

style: fixed 780px panel height — all three equal regardless of content length

style: result body 22px / weight 800 — much bolder display feel

feat: typewriter streaming on AI ANALYZE + bigger bolder body (17px/600) + drop icons

style: equal-height panels (520px min) + Space Grotesk display font + bigger bolder body