
West Lake hand-drawn tour map
Hand-drawn West Lake tour image with preserved SAM-backed hotspots for the lake, causeways, islands, pagoda, hills, and garden regions.
Interactive Visual Images for
Long-Form LLM Answers
Instead of reading a long text block, users can click the parts of an image they care about and ask follow-up questions in context.
Transparent, accurate regions are placed over the generated image so each section can be explored independently.
Each hotspot keeps its own title, summary, and follow-up conversation instead of flattening everything into one chat.
Real-provider mode can use LocateAnything, MiMo-vision, local OCR, and SAM3 refinement to align regions with the rendered image.
Run fully in mock mode, or proxy text, image, and vision providers through the backend so API keys stay off the browser.
Generated ChatImages, hotspots, calibration data, and follow-up threads are stored locally so work survives restarts.
Inspect and correct hotspot bounds when a provider draws content differently from the planned layout.
The hosted showcase is exported from real local runs: the image, hotspot bounds, layout state, and region detail JSON are preserved together.
The user submits a prompt in the ChatImage interface.
A long-form answer is produced by the selected provider or by mock mode.
The answer becomes structured modules, auxiliary modules, and region metadata.
A layout spec assigns each module a region and normalized bounds.
The image provider draws the visual from the structured prompt.
Vision grounding can localize where each module actually appears in the image.
Clicking a hotspot opens its detail panel and a focused follow-up thread.
These curated cases use actual ChatImage generated PNGs and exported hotspot state. Open a demo, click its regions, and reuse the preserved JSON state when needed.

Hand-drawn West Lake tour image with preserved SAM-backed hotspots for the lake, causeways, islands, pagoda, hills, and garden regions.

Strict literary comparison. A split-scene image with clickable poets, landscape, wine-scroll motif, theatre manuscript, and shared legacy region.

Reference-aligned Yuquan campus view with clickable Mao statue, stone pedestal, front lawn, sunset hillside, and framing tree canopies.

Strict academic timeline. Six clickable milestones trace Transformer-based language models from architecture to interactive use.

Fresh daily scene. A breakfast comparison with clickable oatmeal bowl, yogurt cup, sandwich, eggs, fruit, and coffee.

Verified visual scene. A cafe layout with clickable barista, espresso machine, pastry case, window seats, pickup shelf, and queue.

Fresh daily scene. A quiet apartment corner with clickable armchair, bookshelf, floor lamp, window, and tea table.

Fresh daily scene. A record-store corner with clickable listening station, vinyl bins, counter, new arrivals wall, and posters.

Fresh daily scene. A plant-care setup with clickable monstera, watering can, grow light, potting bench, and humidity tray.
No API keys are required for mock mode, so the demo path works immediately after install.
npm install.npm start launches the local server on port 5178.?provider=mock for instant key-free demos..env.example to .env.local, fill in keys, and restart.npm run build when you need static assets in dist/.Clone the project, start mock mode, and inspect an interactive ChatImage locally.