DEV Community

Om Prakash
Om Prakash

Posted on • Originally published at pixelapi.dev

We just launched TunesAPI — Train custom AI models for $0.10 (20x cheaper than FAL.ai)

#ai #api #machinelearning #showdev

We just launched TunesAPI — Train custom AI models for $0.10 (20x cheaper than FAL.ai)

We just shipped something we've been building for months: TunesAPI — a LoRA fine-tuning and inference API that lets developers train custom AI image models on their own data, then generate new images from those models.

The pricing is aggressive by design:

Service TunesAPI FAL.ai Replicate
LoRA Training (SDXL) $0.10 $2.00 ~$3.78
LoRA Training (FLUX) $0.20 $2.00 ~$5.27
Inference (per image) $0.002 $0.035/MP $0.025

That's 20x cheaper for SDXL training. Not a typo.

What does TunesAPI actually do?

It's a 3-step workflow:

Step 1: Train a LoRA on your images

curl -X POST https://api.pixelapi.dev/v1/tunes/train \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "sdxl", "trigger_word": "MYBRAND", "steps": 1000}'
Enter fullscreen mode Exit fullscreen mode

You get back a job ID and an upload URL.

Step 2: Upload 5-100 images of your subject

curl -X POST "https://api.pixelapi.dev/v1/tunes/tune_abc123/upload?token=xyz" \
  -F "images=@photo1.jpg" \
  -F "images=@photo2.jpg" \
  -F "images=@photo3.jpg" \
  -F "images=@photo4.jpg" \
  -F "images=@photo5.jpg"
Enter fullscreen mode Exit fullscreen mode

Training starts automatically. Takes ~15-25 minutes.

Step 3: Generate new images

curl -X POST https://api.pixelapi.dev/v1/tunes/infer \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "tune_id": "lora_xyz",
    "prompt": "a photo of MYBRAND product on a marble table, studio lighting",
    "num_images": 4
  }'
Enter fullscreen mode Exit fullscreen mode

Each image costs 2 credits ($0.002). Done.

Why this matters

LoRA fine-tuning lets you teach an AI model YOUR specific style, product, or subject. Once trained, the model can generate unlimited variations:

  • E-commerce: Train on your product → generate lifestyle shots, different backgrounds, seasonal themes
  • Branding: Train on your visual style → generate consistent brand imagery
  • Real estate: Train on architectural styles → render properties in different aesthetics
  • Fashion: Train on clothing items → generate model photos without photoshoots

The problem was always cost. Astria.ai charges $1.50+ per training. FAL.ai charges $2.00. Replicate charges by GPU time ($3.78+ for a typical FLUX LoRA).

We run on our own GPUs (RTX 6000 Ada, RTX 4070s) — no cloud markup. That's why we can charge $0.10 for SDXL training and $0.002 per inference image.

What's under the hood

  • Training: FLUX.1 Dev and SDXL 1.0 base models, LoRA rank 4-128, 100-5000 steps
  • Inference: Diffusers pipeline with LoRA weights loaded
  • Infrastructure: 5 GPU workers (104GB total VRAM), intelligent job scheduler with priority queuing
  • API: FastAPI, Redis job queue, webhook callbacks for async completion
  • Auth: Same PixelAPI API keys — existing users can start immediately

Who is this for?

Developers building apps that need custom AI image generation. If you're using Replicate, FAL.ai, or Astria.ai and want to cut costs by 10-20x, this is for you.

We also support INR billing and UPI payments for Indian developers — something no global competitor offers.

Try it

  1. Get your free API key: pixelapi.dev/app
  2. You get 100 free credits on signup (enough for 1 training + 50 inference images)
  3. Full docs: pixelapi.dev/docs/tunesapi

Built by PixelAPI — we run our own GPU hardware so you don't pay cloud markup. 14+ AI models, one API.

Top comments (0)