creative_studio

Creative Studio — AI Creative Director

A streaming chat experience where ARIA, an opinionated AI creative director, rewrites your idea into a cinematic prompt and generates the result with DALL-E 3 (images) or Veo 2 (5–8s video clips).

┌─────────────────────────┬──────────────────────────────────────────┐
│  Gallery                │   ARIA — AI Creative Director            │
├─────────────────────────┼──────────────────────────────────────────┤
│  [img] cyberpunk alley  │   > a lone wizard in a storm             │
│  [img] desert dunes     │   ⚙ generate_image (vivid · 1792×1024)  │
│  [vid] dolly-in cafe    │   [image displayed inline]               │
│                         │                                          │
│                         │   I went hyper-real with a low angle …  │
│                         │   Want a 5s slow-zoom video version?    │
└─────────────────────────┴──────────────────────────────────────────┘

Capabilities

Feature	What ARIA can do
🎨 Image generation	DALL-E 3 — landscape / portrait / square, vivid or natural, hd quality
🎬 Video generation	Veo 2 — 5–8 second clips, 16:9 or 9:16, with camera motion
🎯 Creative direction	Rewrites bare ideas into cinematic prompts (camera, lighting, palette, mood)
🖼️ Inline streaming	Media renders in chat the moment the tool returns — no upload step
🗂️ Persistent gallery	Sidebar collects every image/video from the session, click to zoom
💾 Local storage	Files saved to ./generated/ — easy to find, no temp-dir hunting

Quickstart

# 1. API keys (both required for full functionality)
export OPENAI_API_KEY=sk-...        # DALL-E 3
export GEMINI_API_KEY=AIza...       # Veo 2

# 2. Pick your reasoning LLM (defaults to OpenAI)
export LLM_PROVIDER=openai          # or anthropic, gemini

# 3. Run
make example-creative
# or:  cd examples/creative_studio && go run .

# 4. Open
open http://localhost:8890

The UI has a row of spark prompts — click one for an instant demo.

Things to try

Image — let ARIA cook

a wizard in a storm

→ ARIA rewrites this into something like "Low-angle wide shot of a hooded wizard, staff raised against a roiling thunderhead. Cinematic lighting, volumetric rain, painterly Frank Frazetta palette, depth of field." — then calls generate_image with quality=hd.

Video — describe motion, not just a scene

slow dolly-in on a neon Tokyo alley at night

→ Veo 2 renders 5–8 seconds with camera motion baked in. Plays inline.

Compare alternatives

generate three versions of a desert sunset: photorealistic,
oil painting, and anime concept art

→ ARIA fires three separate generations side-by-side in the gallery.

Mix media

make a square image of a rainy cyberpunk diner, then turn it into a
9:16 vertical video clip

→ Image first, then a vertical video version with the same vibe.

Endpoints

Endpoint	Purpose
GET /	Chat UI
GET /api/chat?session_id=&message=	SSE stream — agent turns + media
GET /media/<filename>	Generated images & videos (served from ./generated/)

Configuration

Variable	Purpose	Default
OPENAI_API_KEY	DALL-E 3 image generation	required for images
GEMINI_API_KEY	Veo 2 video generation	required for videos
LLM_PROVIDER	Reasoning LLM: openai, anthropic, gemini	openai
OPENAI_MODEL / ANTHROPIC_MODEL / GEMINI_MODEL	Override model	provider default
MEDIA_DIR	Where to save generated files	./generated/

If OPENAI_API_KEY is set but GEMINI_API_KEY isn't, only generate_image will be registered (and vice-versa). ARIA adapts to whichever tools exist.

Where files land

By default, generated media goes to ./generated/ next to where you ran the command — easy to ls, easy to share. Override with MEDIA_DIR if you want them elsewhere:

MEDIA_DIR=~/Pictures/aria make example-creative

The HTTP server mounts whatever directory you point at as /media/, so URLs in the chat (/media/img-1742...png) always resolve.

How the tools work

pkg/tools/builtin/generate_image.go and generate_video.go:

• Image: calls OpenAI's images.generate API, downloads the CDN URL before it expires (~1h), and saves to SaveDir. Returns markdown ![alt](localURL) so the SSE stream renders it in place.
• Video: calls Veo 2's GenerateVideos, polls the long-running operation every 10s (timeout 5 min), downloads the resulting Gemini Files API URI server-side (the URI requires the API key — the browser can't fetch it directly), and returns an inline <video> tag.

Both tools implement tools.InlineRenderer (InlineResult() bool) so the frontend streams the media into the assistant message immediately, without waiting for the model to format it.

Customizing ARIA

The persona lives in main.go as systemPrompt. Tweak the tone, add new decision rules, or wire in a different toolset — the agent loop handles the rest.