Skip to content
Projects

b3b41020/free-multimodal-proxy

View on GitHub

b3b41020/free-multimodal-proxy has 30 GitHub stars and 25 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-15 21:43 UTC. Last push 2026-08-18.

How these counts are measured

What you can do

  • Text chat — OpenAI-compatible chat/completions, including SSE streaming with reasoning chunks
  • Image generation — 25+ models (Flux, GPT-Image, Seedream, Qwen-Image, Ideogram, Imagen, Wan, ...), returns b64_json
  • Async job API: POST /v1/videos → poll → download MP4
  • Sync API: POST /v1/videos/generations

Measured star growth

in 24h
0
in 7 days
0
in 15 days
Update delayed
in 1 month
Update delayed

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

OpenAI-compatible reverse proxy for free multimodal AI APIs (chat / images / videos / audio / 3d)

Creative & Media

README.md

Original

inferenceport-proxy

English | 简体中文

OpenAI-compatible reverse proxy for InferencePort AI's free multimodal API.

No registration. No API key. Works out of the box with any OpenAI SDK — text chat, image generation, and video generation (sync + async) through one simple endpoint.

What makes this interesting: the upstream backend exposes a public HTTP API that requires no auth token and currently applies no content moderation (equivalent to the app's "studio" mode). This project wraps it in a standard OpenAI-compatible surface so you can use it with your existing tools.

Features

  • Text chat — OpenAI-compatible chat/completions, including SSE streaming with reasoning chunks
  • Image generation — 25+ models (Flux, GPT-Image, Seedream, Qwen-Image, Ideogram, Imagen, Wan, ...), returns b64_json
  • Video generation — 19+ models (Wan, Kling, Vidu, Seedance, PixVerse, Flux, Hailuo, ...)
    • Async job API: POST /v1/videos → poll → download MP4
    • Sync API: POST /v1/videos/generations
  • Audio / 3D passthrough (upstream status dependent)
  • 106 models served from a cached /v1/models listing (5 min TTL)
  • Optional PROXY_TOKEN Bearer auth for public deployments
  • Bounded retries on transient upstream failures (429/5xx/network)

Quick start

git clone https://github.com/b3b41020/free-multimodal-proxy.git
cd free-multimodal-proxy
docker compose up -d --build

That's it. The proxy listens on http://localhost:8080.

# health
curl http://localhost:8080/v1/healthz

# models
curl http://localhost:8080/v1/models

# chat
curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"lightning","messages":[{"role":"user","content":"hello"}]}'

# image
curl http://localhost:8080/v1/images/generations \
  -H 'Content-Type: application/json' \
  -d '{"model":"zimage","prompt":"a cute cat"}' \
  | python3 -c "import json,sys,base64; d=json.load(sys.stdin); open('img.jpg','wb').write(base64.b64decode(d['data'][0]['b64_json']))"

# async video
JOB=$(curl -s http://localhost:8080/v1/videos \
  -H 'Content-Type: application/json' \
  -d '{"model":"wan-fast","prompt":"a dolphin jumping over waves"}' \
  | python3 -c "import json,sys; print(json.load(sys.stdin)['id'])")
curl -s http://localhost:8080/v1/videos/$JOB                # poll: pending -> processing -> completed
curl -s -o video.mp4 http://localhost:8080/v1/videos/$JOB/content   # download

Use with any OpenAI SDK by pointing base_url at http://localhost:8080/v1.

API

Endpoint Description
GET /v1/healthz Health check + upstream reachability
GET /v1/models Upstream model list (cached, 5 min)
POST /v1/chat/completions Chat, stream=true supported (SSE)
POST /v1/images/generations Image generation → data[0].b64_json
POST /v1/videos Create async video job → {id, status}
GET /v1/videos/{id} Poll video job status
GET /v1/videos/{id}/content Download video file (MP4)
POST /v1/videos/generations Sync video generation (can take minutes)
POST /v1/audio/generations Audio generation (upstream dependent)
POST /v1/audio/speech TTS (upstream dependent)
POST /v1/3d/generations 3D generation, requires image_urls (upstream dependent)

Configuration

All settings are environment variables:

Variable Default Description
UPSTREAM_URL https://sharktide-lightning.hf.space Upstream API base
RETRIES 3 Retries for transient upstream failures
CONNECT_TIMEOUT 8 Connect timeout (s)
READ_TIMEOUT 600 Read timeout (s) — sync video can take minutes
CHAT_READ_TIMEOUT 120 Chat read timeout (s)
MODEL_CACHE_TTL 300 Model list cache TTL (s)
PROXY_TOKEN (empty) Optional Bearer token; empty = open access

Popular models

  • Chat: lightning (router), grok-4.5, claude-haiku-4.5, gemini-3.7-flash, gpt-5.6-sol, deepseek-v4-pro, glm-5.2, kimi-k3
  • Image: zimage, lightning-image-turbo, flux-2-pro, gptimage-1.5, seedream-5.0-lite, qwen-image-2.0-pro, ideogram-v4-turbo, imagen-4.0-fast
  • Video: wan-fast, kling-2.1-pro, vidu-q3, seedance-2.5-720p, pixverse-v5, flux-3, hailuo-02

Run GET /v1/models for the full 106-model listing.

Architecture

OpenAI SDK / curl
      │  OpenAI-compatible /v1/* (JSON or SSE)
      ▼
inferenceport-proxy  (FastAPI, Docker)
      │  HTTP /gen/*  (plain, no auth)
      ▼
sharktide-lightning.hf.space  (InferencePort AI free tier)

The proxy is a thin translation layer: /v1/* paths map 1:1 onto the upstream /gen/* API, with model-list caching, timeout handling, and optional auth. No user state is stored.

Deploying elsewhere

  • Docker — docker compose up -d --build (recommended)
  • Bare Python — pip install fastapi uvicorn "httpx[socks]" && uvicorn app:app --host 0.0.0.0 --port 8080
  • Port — change the ports: mapping in docker-compose.yml (e.g. 8098:8080 to expose on 8098)

If your Docker host runs many compose projects and hits "all predefined address pools have been fully subnetted", add a private subnet to the compose file:

networks:
  default:
    ipam:
      config:
        - subnet: 10.99.8.0/24

Disclaimer

  • This is a free tier proxy. The upstream enforces per-identity quotas (chat ~50/day, images ~10/day, videos ~3/day per anonymous identity on the web app); the direct HTTP layer is currently more lenient but may tighten at any time.
  • Audio and 3D endpoints depend on upstream availability and may return errors.
  • Use responsibly and respect the upstream service's terms.
  • This project is an independent open-source implementation and is NOT affiliated with, endorsed by, or sponsored by InferencePort AI or any of its parent companies, subsidiaries, or partners. "InferencePort" is a trademark of its
    respective owner, used here only for identification purposes.

By using this software, you acknowledge and agree that:

  1. Free-tier limitations. The upstream service enforces quotas on its free tier (e.g. per-identity daily limits on chat / image / video generation). These quotas are determined entirely by the upstream and may be tightened, removed, or changed at any time without notice. This project does not and cannot guarantee availability, reliability, or quota levels.

  2. No content review. The upstream API proxied by this software may not apply content moderation. You are solely responsible for the content you generate, transmit, or store using this software, and for ensuring it complies with all
    applicable laws, regulations, and the terms of service of the upstream provider.

  3. Use at your own risk. This software is provided on an "AS IS" and "AS AVAILABLE" basis, without warranties of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular
    purpose, or non-infringement. In no event shall the authors or contributors be liable for any claim, damages, or other liability arising from or in connection with the use of this software.

  4. Compliance. You are responsible for complying with the upstream provider's terms of service and all applicable local laws. Do not use this software for any illegal, abusive, or unauthorized purposes. The upstream may block or
    rate-limit your access at any time; this project is not responsible for any resulting loss.

  5. Upstream dependency. This project depends entirely on a third-party public service that we do not control. The upstream may be discontinued, change its API, or become inaccessible at any time, which would render this software
    non-functional. We make no promises about the long-term viability of this project.

  6. Indemnification. By using this software, you agree to indemnify and hold harmless the authors and contributors from and against any claims, liabilities, damages, losses, or expenses arising out of your use of the software or
    violation of these terms.

For production or commercial use, please refer to the official InferencePort AI service and its licensing terms.

Credits

License

MIT