Fal.ai Models¶

Available models and endpoints used in the Sartiq platform.

Image Generation¶

Flux Models¶

Endpoint	Purpose	Features
`fal-ai/flux/schnell`	Fast generation	Quick iterations
`fal-ai/flux-pro/kontext/max`	High quality with context	Reference images
`workflows/Shootify/flux-krea-lora-i2i`	Custom LoRA workflow	Multiple LoRAs

Generation Parameters¶

{
    "prompt": str,              # Text prompt
    "negative_prompt": str,     # What to avoid
    "image_url": str,           # Reference image (optional)
    "image_size": str,          # "square", "portrait_4_3", etc.
    "num_outputs": int,         # 1-4 images
    "seed": int,                # Reproducibility
    "guidance_scale": float,    # CFG scale (default 7.5)
    "num_inference_steps": int, # Quality vs speed
    "loras": [                  # Custom LoRA models
        {"path": str, "scale": float}
    ]
}

Image Sizes¶

Name	Dimensions	Use Case
`square`	1024x1024	Social media
`portrait_4_3`	768x1024	Product photos
`portrait_16_9`	576x1024	Vertical video
`landscape_4_3`	1024x768	Wide shots
`landscape_16_9`	1024x576	Horizontal video

LoRA Training¶

Flux LoRA Fast Training¶

Endpoint	`fal-ai/flux-lora-fast-training`
Input	ZIP dataset with images
Output	LoRA weights file
Time	~15-30 minutes

{
    "images_data_url": str,     # URL to ZIP dataset
    "steps": int,               # Training steps
    "trigger_word": str,        # Activation word
    "create_masks": bool,       # Auto-mask subjects
    "is_style": bool,           # Style vs subject training
    "is_input_format_already_preprocessed": bool
}

Training Types¶

Type	`is_style`	Use Case
Subject	`False`	Person, product identity
Style	`True`	Visual aesthetic, lighting

Vision & Detection¶

Face Detection (Moondream3)¶

Endpoint	`fal-ai/moondream3-preview/detect`
Input	Image URL + prompt
Output	Bounding boxes

{
    "image_url": str,
    "prompt": "face"  # or other object
}

# Response
{
    "objects": [
        {"bbox": [x1, y1, x2, y2], "confidence": 0.95}
    ]
}

Segmentation (SAM-3)¶

Endpoint	`fal-ai/sam-3/image`
Input	Image URL + text prompt
Output	Segmentation mask

{
    "image_url": str,
    "text_prompt": str,    # e.g., "human person"
    "prompt": str,         # Same as text_prompt
    "apply_mask": bool,    # Apply mask to image
    "output_format": "png"
}

# Response
{
    "masks": [{"url": str, "score": float}]
}

Image Processing¶

Image Upscaling¶

Endpoint	`fal-ai/seedvr/upscale/image`
Scales	2x, 4x, 8x

{
    "image_url": str,
    "scale": 2  # or 4, 8
}

Nano Banana Upscaling¶

Endpoint	`fal-ai/nano-banana-pro/edit`
Use	Quality enhancement

Image Expansion (Outpainting)¶

Endpoint	`fal-ai/bria/expand`
Use	Extend image boundaries

Video Generation¶

Endpoint	`fal-ai/video-gen`
Input	Image + prompt
Output	Video URL

{
    "image_url": str,
    "prompt": str,
    "duration": float,  # seconds
    "fps": int
}

Custom Sartiq Workflows¶

Generation Workflow¶

Endpoint	`workflows/Shootify/flux-krea-lora-i2i`
Features	Multiple LoRAs, reference images

{
    "prompt": str,
    "negative_prompt": str,
    "image_urls": [str],           # Reference images
    "loras": [
        {"path": str, "scale": 0.8}
    ],
    "num_outputs": 4,
    "guidance_scale": 7.5,
    "num_inference_steps": 30
}

Editing Workflow¶

Endpoint	`workflows/Shootify/editing-workflow`
Features	Base image + references

{
    "prompt": str,
    "base_image_url": str,         # Image to edit
    "reference_image_urls": [str], # Style/content refs
    "edit_strength": 0.75          # How much to change
}

Provider Capabilities¶

The Fal integration supports these capability flags:

Capability	Description
`TEXT2IMAGE`	Generate from text
`IMAGE2IMAGE`	Generate from image + text
`IMAGES2IMAGE`	Multiple reference images
`IMAGE2VIDEO`	Generate video from image
`LORAS`	Custom LoRA support
`OUTPUT_SIZE`	Custom dimensions
`UPSCALE_2X/4X/8X`	Fixed scale factors
`INFERENCE_STEPS`	Control quality
`NEGATIVE_PROMPT`	Exclusion prompts
`INFERENCE_GUIDANCE`	CFG scale
`EDIT_STRENGTH`	Edit intensity
`OUTPUT_DURATION`	Video length