fal-realtime

Real-time and streaming image generation using fal.ai's fastest models. Results in under 1 second — ideal for interactive applications, live previews, and rapid iteration.

Scripts

Script	Purpose
realtime.sh	Generate images in real-time (sub-second)

Usage

Quick Real-Time Generation

./scripts/realtime.sh --prompt "A cute cat wearing a top hat"

With LoRA

./scripts/realtime.sh --prompt "A portrait in sks style" --lora-url "https://example.com/lora.safetensors"

Multiple Rapid Iterations

./scripts/realtime.sh --prompt "Abstract art with flowing colors, version 1"
./scripts/realtime.sh --prompt "Abstract art with flowing colors, version 2, more vibrant"
./scripts/realtime.sh --prompt "Abstract art with flowing colors, version 3, darker tones"

Arguments

Argument	Description	Required
--prompt / -p	Image description	Yes
--model / -m	Model endpoint	No (default: fal-ai/flux-2/klein/realtime)
--size	square, landscape, portrait	No (default: square)
--seed	Random seed for reproducibility	No
--lora-url	URL to LoRA weights	No
--lora-scale	LoRA influence (0.0-1.0, default: 1.0)	No
--num-images	Number of images (default: 1)	No

Finding Models

To discover the best and latest real-time generation models, use the search API:

# Search for real-time / fast image generation models
bash /mnt/skills/user/fal-generate/scripts/search-models.sh --query "realtime"
bash /mnt/skills/user/fal-generate/scripts/search-models.sh --query "fast image"
bash /mnt/skills/user/fal-generate/scripts/search-models.sh --category "text-to-image"

Or use the search_models MCP tool with keywords like "realtime", "fast", "schnell", "turbo".

When to Use Real-Time vs Standard Generation

Use Case	Use Real-Time	Use Standard
Rapid prototyping / iteration	Yes
Interactive apps / live preview	Yes
Final high-quality output		Yes (FLUX Dev, Nano Banana Pro)
Professional / commercial work		Yes (FLUX Pro Ultra)
Exploring prompt ideas quickly	Yes
Batch generation with variations	Yes

Output Format

{
  "images": [
    {
      "url": "https://fal.media/files/...",
      "content_type": "image/jpeg",
      "width": 1024,
      "height": 1024
    }
  ],
  "seed": 12345,
  "has_nsfw_concepts": [false]
}

Tips

• Real-time models use fewer inference steps — prompts should be clear and concise
• Use --seed to lock in a good composition, then iterate on the prompt
• LoRA support works with FLUX Klein Realtime — great for style-locked fast iteration
• These models use the synchronous API (no queue), so results are instant