⚡ Qwen 2.1 Turbo + SeedVR2 4K Studio

Self-Hosted, Production-Grade AI Creative Suite: Lightning 8-Step Text-to-Image, Vision-Language Photo Editing, and Cinema 4K Neural Super-Resolution.

Key Features • Web Interface • FinOps Benchmarks • Architecture • Deployment • REST API

---

💡 What is Qwen Upscaler Studio?

Qwen Upscaler Studio is an all-in-one open-source production stack that unifies state-of-the-art vision and diffusion models into an optimized containerized workflow:

1. Qwen 2.1 Turbo (NVFP4 Quantization): Distilled 8-step lightning text-to-image generator producing photorealistic anatomical accuracy, complex lighting, and typography in just seconds on any modern GPU.
2. Qwen3-VL 8B Reference Conditioning: Multimodal instruction-guided image editing. Modify clothing, backgrounds, or cinematic lighting while strictly preserving subject facial identity, pose, and physical proportions.
3. SeedVR2 (EMA 3B FP8): State-of-the-art temporal diffusion super-resolution model for upscaling photography and video up to 4K UHD (2160p 60fps) with perceptual reconstruction and color matching.
4. 1-Click 4K Combo Pipeline: Generate a fresh concept with Qwen Turbo and upscale it directly into fine 4K micro-textures through an automated orchestration pipeline with intelligent VRAM offloading.

---

✨ Key Features

• ⚡ Instant 8-Step Synthesis: Photorealistic generation in 8 distilled steps, saving up to 90% in cloud GPU compute costs compared to traditional 30-50 step diffusion models across any modern NVIDIA GPU.
• 🔍 4K Photo & Video Super-Resolution: Upscale static imagery or 60fps videos using temporal window tiling, FP8 attention, and customized FFmpeg NVENC acceleration.
• 🔄 Zero-OOM VRAM Orchestrator: Runs both ComfyUI diffusion graphs and SeedVR2 super-resolution inside a single 24GB VRAM envelope via dynamic memory flushing (free_comfyui_memory() + PyTorch CUDA cache clearing).
• 📱 Fully Responsive Modern UI: Built with Tailwind CSS, dark glassmorphism, prompt presets, seed controls, real-time GPU telemetry, and an outputs gallery. Adapts seamlessly to mobile viewports.
• ☁️ Multi-Cloud Native: 1-click turnkey deployment on SaladCloud, Vast.ai, local NVIDIA workstations, or as a standalone Hugging Face Space.
• 🔒 Zero-Setup Container: All weights, compiled CUDA kernels, dependencies, and web services are pre-baked into s3rfdev/qwen-turbo-upscaler:latest. Just run and generate!

---

🖼️ Live Web Interface & Gallery

1-Click 4K Combo Pipeline	SeedVR2 4K Super-Resolution

Mobile Viewport (Responsive)	Sample Photorealistic Output

---

📊 FinOps & Hardware Benchmarks

Cost per 1,000 Images Across Cloud GPUs

GPU Model	VRAM	Architecture	Hourly Rate (SaladCloud)	Inference Time (T2I 1024²)	Cost / 1,000 Images	Cost Efficiency Rating
NVIDIA RTX A5000	24 GB	Ampere	$0.18 / hr	3.2s	$0.160	🏆 Best Value (Cost Leader)
NVIDIA RTX 3090 Ti	24 GB	Ampere	$0.19 / hr	3.0s	$0.158	🥈 Excellent High-Throughput
NVIDIA RTX 4090	24 GB	Ada Lovelace	$0.34 / hr	2.5s	$0.236	⚡ Fast Turnaround
NVIDIA RTX 5090	32 GB	Blackwell	$0.48 / hr	1.7s	$0.226	🚀 Speed Champion (Sub-2s)

Tip

Why Cloud Desktop GPUs Beat High-End Laptops: Mobile laptop GPUs (e.g., RTX 4090 Mobile) are constrained by tight thermal limits (115W–175W TGP) and narrow 16GB VRAM buses, causing thermal throttling and memory offloading during 4K super-resolution passes. Cloud desktop cards (RTX A5000 / 3090 Ti / 4090) operate at 300W–450W with 24GB+ unconstrained GDDR6X/ECC memory, providing consistent, sustained high-throughput generation.

---

🏗️ Architecture

flowchart TD
    Client["Web Client / HF Space / Mobile"] -->|HTTP / REST API| Gateway["FastAPI Server (:8000)"]
    Gateway --> Queue["In-Memory Task Queue & State Store"]
    Queue --> Worker["Asynchronous Pipeline Worker"]
    
    subgraph Execution Pipeline ["GPU Memory & Model Orchestrator"]
        Worker -->|Step 1: Text-to-Image / Editing| Comfy["ComfyUI Headless Engine<br/>• Qwen 2.1 Turbo (NVFP4)<br/>• Qwen3-VL 8B Reference"]
        Comfy --> Flush["VRAM Offload Manager<br/>free_comfyui_memory() + torch.cuda.empty_cache()"]
        Flush -->|Step 2: 4K Super-Resolution| SeedVR["SeedVR2 Inference Engine<br/>• EMA 3B FP8 Diffusion<br/>• FFmpeg NVENC (x264/x265)"]
    end
    
    SeedVR --> Storage["Output Storage & Gallery Cache"]
    Storage --> Client

Loading

---

🚀 Quick Start & Deployment

Option A: 1-Click SaladCloud Deployment (Recommended)

SaladCloud offers enterprise-grade consumer and workstation GPUs at wholesale spot prices (~$0.18/hr).

1. Clone the repository:

   git clone https://github.com/s3rfdev/qwen_upscaler.git
   cd qwen_upscaler

2. Configure your API keys:

   cp .env.example .env

   Add your SALAD_API_KEY and SALAD_ORGANIZATION_NAME from Salad Portal.

3. Deploy using the included CLI:

   # Deploy on cost-effective RTX A5000 24GB
   python ctl.py salad up --gpu rtx_a5000
   
   # Check status and retrieve public endpoint
   python ctl.py salad status

4. Once running, open the URL returned by python ctl.py salad status in your browser.

---

Option B: Local Workstation (Docker Compose)

For local development on an NVIDIA GPU (24GB+ VRAM recommended):

docker compose up -d

Open http://localhost:8000 in your web browser.

Alternatively, with Docker CLI:

docker run --gpus all \
  -d --restart unless-stopped \
  -p 8000:8000 \
  --name qwen-turbo-upscaler \
  s3rfdev/qwen-turbo-upscaler:latest

---

Option C: Vast.ai Deployment

1. Install Vast CLI: pip install vastai
2. Search and create an instance:

   vastai create instance <OFFER_ID> \
     --image s3rfdev/qwen-turbo-upscaler:latest \
     --disk 80 \
     --env '-p 8000:8000 -e PORT=8000' \
     --onstart-cmd "/app/start.sh"

---

Option D: Hugging Face Space Connection (Web Client)

A live web client is hosted at huggingface.co/spaces/s3rfdev/qwen_upscaler.

Note

Why the Space requires a GPU backend: Generating 4K imagery and neural super-resolution requires a 24GB+ GPU. To keep public hosting 100% free without incurring Hugging Face's expensive $1.05/hr GPU billing, this Space operates as a decoupled frontend client.

1. Interactive Demo Mode: Click Try Demo Mode in the header banner to explore the interface, inspect gallery outputs, and simulate generations.
2. Live GPU Generation: Click Endpoint in the header, paste your deployed SaladCloud, Vast.ai, or Local Docker address (e.g. https://your-node.salad.net), and start generating instantly.

---

🛠️ Configuration (.env)

Variable	Description	Default
SALAD_API_KEY	Your SaladCloud API access token	None
SALAD_ORGANIZATION_NAME	Your SaladCloud organization name	None
SALAD_PROJECT_NAME	SaladCloud project identifier	default
SALAD_CONTAINER_NAME	Container group name	qwen-turbo-upscaler
SALAD_GPU_CLASS	Target hardware (rtx_a5000, rtx_3090_ti, rtx_4090, rtx_5090)	rtx_a5000
HF_TOKEN	Hugging Face User Access Token (Write)	None
BACKEND_API_URL	Remote API URL for frontend connection	http://localhost:8000

---

📡 REST API Reference

The container exposes a standard FastAPI backend for easy programmatic integration:

1. Text-to-Image Generation

curl -X POST http://localhost:8000/api/generate \
  -F "prompt=Futuristic cyberpunk architecture, neon reflections, rain, 85mm cinematic photo" \
  -F "width=1024" \
  -F "height=1024" \
  -F "seed=42"

Response:

{
  "job_id": "gen_1791640822",
  "status": "queued",
  "message": "Generation job submitted successfully"
}

2. Instruction-Guided Image Editing

curl -X POST http://localhost:8000/api/edit \
  -F "file=@portrait.png" \
  -F "prompt=Keep the exact person and pose. Change dress color to royal blue satin."

3. SeedVR2 4K Super-Resolution

curl -X POST http://localhost:8000/api/upscale \
  -F "file=@image.png" \
  -F "media_type=image" \
  -F "resolution=2160"

4. 1-Click 4K Combo Pipeline

curl -X POST http://localhost:8000/api/combo \
  -F "prompt=A luxury fashion editorial photograph, ivory silk gown, brutalist architecture" \
  -F "width=1024" \
  -F "height=1024"

5. Telemetry & GPU Stats

curl http://localhost:8000/api/stats

Response:

{
  "gpu": {
    "name": "NVIDIA RTX A5000",
    "total_gb": 24.0,
    "free_gb": 21.8,
    "allocated_gb": 2.2
  },
  "queue_count": 0,
  "active_jobs": 0
}

---

🤝 Contributing & License

Contributions, issues, and feature requests are welcome!

Distributed under the Apache 2.0 License. See LICENSE for details.

Developed with ❤️ by s3rfdev.