The terms behind every tool on ImageLabs, defined in one or two honest sentences.
Alpha channel: The per-pixel transparency channel in a PNG, WebP or AVIF file. An alpha channel stores how opaque each pixel is, from fully transparent to fully solid. Partial alpha is what makes hair edges, glass and soft shadows composite cleanly. JPG has no alpha channel, which is why a 'transparent JPG' does not exist.
Matting: Predicting a continuous transparency value per pixel, rather than a hard cutout. Segmentation labels each pixel subject or background. Matting predicts a value in between, which is the only way to represent a pixel that is genuinely half hair and half background. Models like BiRefNet are matting models, which is why they handle fur, glass and motion blur.
BiRefNet: A high-fidelity matting network used for strand-level background removal. BiRefNet (Bilateral Reference Network) is a dichotomous image segmentation model designed for high-resolution, fine-detail subjects. It is the engine ImageLabs uses when a cutout involves hair, fur, foliage or transparency.
Super-resolution: Enlarging an image by predicting detail, not by interpolating pixels. A super-resolution network is trained on pairs of low- and high-resolution images and learns what plausible detail looks like at a larger size. This is why AI upscaling recovers texture where bicubic resampling only softens.
Real-ESRGAN: The widely used practical super-resolution model behind most AI upscalers. Real-ESRGAN extends ESRGAN with training on synthetically degraded images, so it handles real-world compression artifacts and noise rather than only clean downsamples. ImageLabs uses it for 2x and 4K upscaling.
GFPGAN: A face-specific restoration model that rebuilds degraded facial detail. General upscalers smear faces because faces carry structure a generic model does not know about. GFPGAN uses a facial prior to restore eyes, mouths and skin texture while keeping identity recognisable.
CodeFormer: A face restoration model with a fidelity/quality trade-off control. CodeFormer restores faces through a learned codebook, and exposes a weighting between staying faithful to the original pixels and producing a cleaner face. It is the default one-credit restoration engine on ImageLabs.
Inpainting: Regenerating a masked region of an image from its surroundings. Inpainting fills a selected area with content that matches the rest of the image — used for removing objects, watermarks and blemishes. Unlike clone-stamping, it synthesises new texture rather than copying existing pixels.
Outpainting: Extending an image beyond its original frame. Outpainting is inpainting applied outside the canvas: the model invents what would have been in frame if the shot were wider. Useful for changing aspect ratio without cropping the subject.
FLUX Fill: A diffusion model specialised for inpainting and outpainting. FLUX Fill is a variant of the FLUX family trained specifically for masked-region generation, which makes its fills blend with surrounding lighting and texture better than a general text-to-image model prompted to inpaint.
Nano Banana Pro: Google's Gemini 3 Pro Image model, used for generation and instruction-based editing. Nano Banana Pro follows natural-language editing instructions on an existing image — change a garment, a pose, a background — while preserving identity and lighting. It also generates images from text and is notably good at legible text inside images.
Diffusion model: A generative model that turns noise into an image through iterative denoising. Diffusion models learn to reverse a gradual noising process. Starting from random noise and a conditioning signal such as a text prompt, they denoise step by step until an image emerges. Most current image generators are diffusion or diffusion-transformer based.
Prompt: The text instruction that conditions a generation or edit. For editing, the most effective prompts are physical and specific: name the object, the change and what must stay the same. For generation, subject plus setting plus lighting plus lens gives more control than a long adjective list.
Seed: The random starting point that makes a generation reproducible. The same prompt with the same seed and settings produces the same image. Changing only the seed gives a different interpretation of the same prompt, which is the standard way to explore variations.
Aspect ratio: The width-to-height proportion of an image, such as 16:9 or 1:1. Generators produce images at fixed ratios rather than arbitrary pixel sizes. Choosing the ratio up front avoids cropping later: 16:9 for thumbnails and banners, 1:1 for avatars and listings, 9:16 for stories and reels.
Upscaling vs resampling: Resampling interpolates existing pixels; upscaling predicts new detail. Bicubic or Lanczos resampling averages neighbouring pixels — the result is larger and softer. AI upscaling infers detail that was never captured, which is why texture reappears rather than blurring.
Lossy vs lossless compression: Lossy formats discard detail permanently; lossless formats do not. JPG, WebP (lossy) and AVIF trade fidelity for size and cannot be fully reversed. PNG and lossless WebP keep every pixel. Re-saving a lossy file repeatedly compounds artifacts, so keep an original master.
Vectorization: Tracing a raster image into resolution-independent paths. Vectorization converts regions of colour into filled paths that scale infinitely and stay editable. A genuine SVG contains paths; a fake one contains an embedded bitmap and pixelates exactly like the original.
VTracer: An open-source colour tracing engine that converts raster art to SVG. VTracer clusters pixels into colour regions and fits curves to their boundaries, handling both flat logo art and photographic gradients. It is the engine behind ImageLabs' vectorizer.
Denoising: Removing sensor noise and grain without erasing real detail. High-ISO and night photos carry luminance and colour noise. A learned denoiser separates noise from texture, unlike a blur which removes both. Denoise before sharpening or upscaling, never after.
Colorization: Adding plausible colour to a black-and-white image. Colorization is inference, not recovery: the original colour information does not exist in a greyscale negative. Models apply learned priors for skin, sky, foliage and materials, which is reliable in general and speculative for specific dyed or painted objects.
Identity preservation: Keeping a person recognisably themselves through an AI edit. The difference between editing a photo and regenerating one. Instruction-based editors constrain changes to the region described so the face, hands and proportions stay as photographed — the requirement for headshots, outfit swaps and pose changes.