Key Concepts

Key Concepts

Models

Model is a type of machine learning model that is trained to generate new data, such as text, images, audio, or video, based on the patterns and relationships it has learned from a training dataset. To see the list of all public models on Segmind, go to segmind.com/models and filter by models.

Segmind offers a combination of third-party models, accessed through direct integrations with partner APIs, and self-hosted models, which are deployed and managed on GPUs controlled by Segmind across various cloud providers.

Latency

Segmind is designed to deliver best-in-class inference latency, powered by a highly optimized serving infrastructure. Our inference engine ranks among the fastest globally, ensuring quick and reliable generation across all supported models.

The latency metrics displayed on each model’s page represent average values across all users. Actual latency for your requests may vary depending on factors such as input parameters, the GPU type used, and real-time server load.

For billing, Segmind measures the exact GPU processing time used to execute your request, ensuring you only pay for the actual compute resources consumed.

PixelFlow GUI

Pixelflow is a node based tool that gives developer and creators the ability to access a myriad of open-sourced models and seamlessly string them together to create highly tailored AI workflows.

PixelFlow APIs

Workflows created on PixelFlow GUI can be converted into APIs for simpler calling from your application. You can mark the inputs and outputs to a workflows to convert them into PixelFlow APIs. See the PixelFlow API reference for the request shape and how to call a published workflow.

Playground

Playgrounds allow you to use a model or a PixelFlow via a simple user interface. This is a simple form interface to give in the inputs and view the outputs without using an API key.

Fine-tuning

Fine-tuning refers to taking a pre-trained image model and training it further on a specific dataset, so it specialises for a particular subject, style or task.

Segmind's fine-tuning service covered the Flux family and Qwen. It no longer accepts new training jobs — existing fine-tuned models remain available for inference. See the fine-tuning overview.

Prompts

Prompts are textual inputs or instructions provided to generative AI models like text-to-image, image-to-image, and language models. For text-to-image models, prompts describe the desired visual content, guiding the generation of images. In image-to-image models, prompts specify the desired modifications or transformations to an existing image. Language models use prompts as context or starting points for generating human-like text.

Effective prompting is crucial, as the quality and specificity of prompts significantly influence the model's output. Prompt engineering involves carefully crafting prompts to elicit the desired responses, often requiring experimentation and iterative refinement for optimal results.

Tokens

Tokens are the basic units of text that large language models (LLMs) process and generate, typically subword units like words, punctuation, or word pieces. LLMs are charged per token because tokenizing inputs, generating outputs, and the overall computational cost scale with the number of tokens.

Longer inputs/outputs and larger models require processing more tokens, consuming more resources. Charging per token allows providers to account for the variable computational demands based on the input/output length and model size for each request. It creates a scalable pricing model tied to the actual resources consumed for running these complex models.

Weights

Weights refer to the numerical parameters of the deep neural network that encode the mapping between text prompts and generated images. LoRA (Low-Rank Adaptation) and other fine-tuning techniques aim to specialize or adapt these weights for specific domains or tasks without modifying the entire pre-trained model. LoRA adds a small set of task-specific weights on top of the base model weights during fine-tuning. This allows efficient customization while preserving the general knowledge from the original training. Other techniques like prompt tuning or full fine-tuning update more weights for increased specialization. Ultimately, these methods optimize a subset of weights to encode new knowledge while leveraging the robustness of the pre-trained weights, enabling efficient domain adaptation of large generative models.

Base model

A base model for a weight file refers to the initial, pre-trained model that serves as the foundation for further fine-tuning or adaptation through techniques like LoRA or prompt tuning.

Modality

Modality refers to the different types or forms of data that a system can process. Common modalities include: Text, Image, Audio and Video. Segmind supports the following model types: