Standalone, high-performance Generative AI creative suite powered by Bun, Higgsfield AI, and local FFmpeg (Storage Architecture & Multi-Format Export Suite).
Quick Start β’ DevOps Architecture β’ Benchmarks β’ AGENTS.md β’ Creators
VEFA: Creator Studio v0.5.0 is an isolated, edge-accelerated generative media workstation and orchestration platform. It unifies cloud multi-model generative AI (Higgsfield AI, Seedance 2.5, Kling 1.5, FLUX.2, GPT Image 2.5, Seed Audio 1.0) with local FFmpeg Non-Linear Video Editing (NLE), hierarchical date-partitioned workspace storage (output/{workspace_id}/{YYYY-MM-DD}/), SHA-256 integrity checksums, multi-format social/production export presets (reels_9_16, cinematic_16_9, animated_webp, animated_gif, social_preview), and zero-dependency Cloudflare R2 / AWS S3 SigV4 cloud sync.
Engineered for zero-terminal onboarding, creators synthesize cinematic soundscapes, trim video sequences, export 1-click social presets, and back up assets to cloud buckets directly in the browser without executing command-line scripts.
Benchmarked on Bun 1.4.2 runtime:
| Metric / Benchmark | v0.4.0 | v0.5.0 | DevOps Impact & Technical Notes |
|---|---|---|---|
| HTTP + WebSocket Cold Boot | Instant process boot via native Bun.serve
|
||
| SHA-256 Integrity Checksum | N/A | High-speed cryptographic SHA-256 hashing | |
| AWS SigV4 Header Generation | N/A | Pure Bun primitives (zero bulky AWS SDKs) | |
| Reels (9:16) Blurred Pillarbox | N/A | Smart background blur transcode for TikTok/Reels | |
| Animated WebP / GIF Transcode | N/A | 2-pass palettegen & lanczos web loop exports | |
| Media Metadata Probing | Native regex probing via FFmpeg banner (zero ffprobe dependency) | ||
| Precise Clip In/Out Trimming | Fast stream-copy with sample-accurate transcode fallback | ||
| NLE Transition Rendering | Resilient xfade + fade/overlay dual-mode filtergraph |
||
| Sequential Concat (2 clips) | Zero-re-encode lossless stream copy (-c copy) |
||
| FTS5 Prompt Search (10k items) | C-speed SQLite full-text search index queries | ||
| Single-Binary Compilation | Self-contained executable with embedded runtime |
The entire studio executes within the native Bun runtime. It maintains an inviolable directive: strictly zero host-shell dependencies (no PowerShell, CMD, or Bash invocations). All cross-platform sub-process commands are orchestrated using the built-in Bun Shell ($):
import { $ } from "bun";
const result = await $`bunx @higgsfield/cli ${args}`.cwd(OUTPUT_DIR).quiet().nothrow();- HTTP Engine: Backed by
Bun.servehandling concurrent static asset routing and REST endpoints. - WebSocket Pub/Sub: Real-time terminal telemetry, generation progress, queue updates, and auth state broadcasted via native
server.publish("studio-telemetry", ...). Zero third-party message brokers required.
- Zero Native Compilation: Employs built-in
bun:sqlite(Databaseclass) with zeronode-gypor external C bindings. - Concurrency: Operates in WAL (Write-Ahead Logging) mode (
PRAGMA journal_mode = WAL;) for concurrent reads during active background writes. - FTS5 Full-Text Search: Prompts and metadata are indexed in an SQLite
FTS5virtual table, enabling instant search as you type.
- Headless & Remote Ready: When running on headless servers, containers, or remote developer workspaces, the studio detects loopback approval URLs and provides an instant "Copy Link" button.
- Dynamic Port Negotiation: Prevents port conflicts by scanning ports sequentially starting at
8765. - Fault-Tolerant Cleanup: 5-minute timeout guards terminate stale login processes, resetting session state cleanly.
- Higgsfield AI safety filters or upstream model timeouts occasionally refund credits with messages like
"credits returned"or"credits refunded". HiggsfieldService.parseErrorOutputintercepts these patterns, prevents fatal crashes, alerts the user with a reassuring toast, and synchronizes the live header credit balance badge.
Prerequisites: Bun
# 1. Clone repository
git clone https://github.com/VEFAorg/vefa-creator-studio.git
cd vefa-creator-studio
# 2. Install dependencies
bun install
# 3. Launch studio in development mode
bun run devOpen your browser at http://localhost:3000.
Zero dependencies required. Download the compiled binary for your operating system:
| Platform | Architecture | Binary Asset | Checksum (SHA-256) |
|---|---|---|---|
| Windows | x64 | vefa-creator-studio-windows-x64.exe |
1daa09141b1f15f667ecb9b91adf52a7be0190fca36b0b93e9e549b134a0597a |
| Linux | x64 | vefa-creator-studio-linux-x64 |
12939608339b5c8d7e4c27e17b9594541f1f830027395f7ea8cced79639ff637 |
| macOS | Apple Silicon | vefa-creator-studio-darwin-arm64 |
6e099c9a4c92dc40ae177ca616d03b7cf63759d16b13e1c02170c99294b0d40c |
Build your own binary anytime in
bun run build:binary# Build multi-stage production container
docker build -t vefa-creator-studio:v0.5.0 .
# Run container on port 3000
docker run -d -p 3000:3000 --name creator-studio vefa-creator-studio:v0.5.0To ensure frictionless repository uploads and avoid GitHub web interface rejections:
- Strict File Limit: Total source repository files are maintained under 35 files (well below GitHub's 100-file web upload cap).
- Zero Hidden Dot-Folders: No
.agents/,.bun-cache/, or runtime caches are committed to Git. - Dynamic Skills Acquisition: Higgsfield companion skills are downloaded dynamically in the browser via
POST /api/skills/installinto local application storage.
Execute all 9 test suites:
bun testbun test v1.4.2
β test/auth.test.ts (Port negotiation, loopback lifecycle)
β test/refund.test.ts (Credit refund regex matching & friendly translation)
β test/ffmpeg.test.ts (Lossless stream copy concat & poster extraction)
β test/nle_timeline.test.ts (In/Out trimming, visual transitions, BGM mixing & ducking)
β test/export_storage.test.ts (Hierarchical buckets, SHA-256, 5 export presets, AWS SigV4)
β test/database.test.ts (bun:sqlite schema, WAL mode, FTS5 full-text search)
β test/queue.test.ts (Dynamic worker concurrency auto-tuning)
β test/skills.test.ts (Companion skills status & background downloader)
β test/server.test.ts (HiggsfieldService methods & catalog endpoints)
9 of 9 test suites passed (100% pass rate)
GET /api/status: System health, authentication state, and detected plan tier.POST /api/auth/start: Initiates loopback OAuth login with dynamic port discovery.POST /api/auth/cancel: Terminates pending authentication loopback processes.GET /api/account/balance: Fetches real-time credit balance and plan details.POST /api/generate: Enqueues generative job (imageorvideo) with auto-tuned concurrency.POST /api/stitch: Executes local FFmpeg video concat demuxing.POST /api/export: Transcodes media into presets (reels_9_16,cinematic_16_9,animated_webp,animated_gif,social_preview).GET /api/sync/status: Retrieves Cloudflare R2 / AWS S3 backup status.POST /api/sync/upload: Uploads local creation to Cloudflare R2 / AWS S3 using native SigV4.GET /api/gallery: Lists creations with optional filtering (image,video,stitched,favorites).GET /api/gallery/search?q=: Instant SQLite FTS5 full-text prompt search.POST /api/skills/install: Background download and setup of companion skills.
job_started/job_completed: Live generation state updates.stitch_started/stitch_completed: NLE timeline rendering progress.export_started/export_completed: Multi-format preset transcode updates.cloud_sync_completed: Instant telemetry on cloud bucket uploads.job_refunded: Friendly alert broadcast when upstream model cancels and refunds credits.auth_started/auth_url_ready/auth_success: Non-blocking OAuth progress updates.concurrency_changed: Telemetry when worker queue auto-tunes parallel capacity.
- VEFAorg β Architectural vision, product requirements, and generative ecosystem coordination.
- Devin Damon Shinkle β Lead conception, system design, and engineering leadership.
- Antigravity (Google DeepMind) β Autonomous agentic systems architecture, Bun native optimization, and test automation.
This project is licensed under the MIT License β see the LICENSE file for details.