The server is anavigator.Your browser is the renderer.
CoreAgents.app is a CDN-native, stateless, decentralized image synthesis platform. We use Nemotron-3-Ultra-550B only to optimize prompts into JSON specs. The heavy diffusion compute runs on your GPU via WebGPU. Vercel Edge CDN caches every response for a year — zero function invocations, zero API cost, zero rate limits.
Nine steps. Zero pixels cross our servers.
Every other image-gen API is a server-heavy pipeline: prompt in → GPU farm → image out → bandwidth bill. CoreAgents inverts this. The server is a thin navigation layer cached at the edge; the user's browser does all the heavy lifting via WebGPU.
GET /api/v1/navigate
Client sends a single GET request with ?prompt=...&style=...&aspect_ratio=...&seed=...
Edge CDN Lookup
Vercel's Edge Network hashes the normalized query string. Equivalent prompts collapse to the same cache entry.
Cache HIT → instant JSON
If cached, return immediately. ZERO function invocations, ZERO Nemotron calls, ZERO latency.
Cache MISS → Edge Function
On miss, the Edge Function calls Nemotron-3-Ultra-550B to optimize the prompt into a compact JSON spec.
Nemotron Navigator
Nemotron returns only {optimized_prompt, palette, tags, seed, guidance, steps}. NO pixels. NO heavy compute.
CDN Caches Forever
Response ships with Cache-Control: s-maxage=31536000. Cached at 18 edge regions for a year.
Browser Downloads Model
Browser fetches quantized model directly from HuggingFace CDN. Vercel bandwidth consumed: 0 bytes.
WebGPU Synthesis
Browser runs multi-step diffusion locally on the user's GPU. 4-step turbo takes ~600ms.
Canvas Composite
Final latent grid composited to <canvas>. Serverless from here on. The image lives only in the user's browser.
Why this wins · CoreAgents vs. everyone else
| Dimension | Traditional API (e.g. OpenAI Images) | CoreAgents.app |
|---|---|---|
| Server compute | GPU farm, $$/image | $0.00 — browser GPU |
| Per-image cost | $0.02 – $0.17 | $0.00 after first cache miss |
| Rate limit | 5 – 50 req/min | Unlimited (CDN absorbs) |
| Auth required | API key, billing | None |
| Latency (warm) | 2 – 8 seconds | <50ms cache hit |
| DDoS resistance | Rate limiter, WAF | CDN IS the defense |
| Bandwidth | Server pays egress | HuggingFace pays egress |
| Horizontal scale | Provision GPUs | Infinite (it's just JSON) |
Nemotron-3-Ultra-550B doesn't draw. It navigates.
Running a 550-billion-parameter model on every image request would bankrupt any operator. CoreAgents uses Nemotron exclusively as a latent-space navigator: it receives your prompt and returns only a compact JSON spec — expanded tags, palette, composition hints, lighting direction, and a deterministic embedding seed. The actual pixel synthesis happens on your GPU, not ours.
Because the navigator output is deterministic and CDN-cached, we only pay for the first request for any given canonical prompt. Every subsequent request — from anyone, anywhere — hits the edge and costs nothing. We default to OpenRouter's free tier (nvidia/nemotron-3-ultra-550b-a55b:free, 1M-token context, $0), so the navigator itself runs at literally zero cost.
Nemotron rewrites your prompt into a richer tag list semantically aligned with the style preset.
Automatically generated to suppress common failure modes (blur, artifacts, extra limbs).
Returns 4–6 hex colors tuned to the style. The WebGPU shader quantizes output to this palette.
Drives focal-point placement and key/rim/fill light angles in the WGSL compute shader.
"a lonely lighthouse on a stormy cliff at dusk"output: JSON spec only — no pixels
{
"optimized_prompt": "lone lighthouse,
storm cliff, cyberpunk neon…",
"palette": ["#ff2d92","#9eff00",
"#0a0a0a","#ffd400"],
"composition_tags": [
"rule-of-thirds",
"focal-asymmetry"
],
"lighting_tags": [
"rim-light","key-light-45deg"
],
"seed": 3174268491,
"guidance": 7.5,
"steps": 4
}compute: local GPU · 4-step diffusion · ~600ms
A single GET endpoint. No auth. No SDK.
The entire public API is one URL. Every parameter lives in the query string, so requests are cacheable by every CDN on earth. Curl it, fetch it, drop it into a Python script — no keys, no billing, no rate limits.
Returns a JSON navigation spec for the browser-side WebGPU engine. Cached at the edge for one year — equivalent prompts collapse to the same cache entry automatically.
Parameters
promptstringrequiredThe natural-language prompt. Normalized: lowercased, articles removed, tokens sorted. Equivalent prompts share a cache entry.
styleenumdefault: photorealOne of: photoreal, anime, cyberpunk, low-poly, watercolor, oil, ink, vaporwave, brutalism, noir.
aspect_ratioenumdefault: 1:1One of: 1:1, 16:9, 9:16, 4:3, 3:4, 21:9, 3:2, 2:3.
stepsintegerdefault: 4Diffusion steps (1–8). Lower = faster, higher = sharper. 4 is the turbo sweet spot.
guidancefloatdefault: 7.5Classifier-free guidance scale (0.1–12.0). Higher = more prompt-adherent, lower = more creative.
seedintegerdefault: autoDeterministic seed. If omitted, derived from the prompt hash — same prompt = same seed.
modelenumdefault: flux-schnell-w4Quantized model id: sd-turbo-w4, sdxl-turbo-w4, flux-schnell-w4, kandinsky-3-w4.
# Basic call — returns the navigator JSON spec
curl "https://coreagents.app/api/v1/navigate?prompt=a%20lonely%20lighthouse%20on%20a%20stormy%20cliff&style=cyberpunk&aspect_ratio=16:9"
# Response headers (note the CDN cache directives):
# X-CoreAgents-Hash: a3f7b2c1
# X-CoreAgents-Cache: MISS
# Cache-Control: public, max-age=86400, s-maxage=31536000, immutable
# Repeat the same call → X-CoreAgents-Cache: HIT, served from edge in <50msTry it
DMCA
著作権に関する問い合わせ窓口と対応ポリシーをこのセクションに整理します。