Internal Links

• Execution Tools (Image, LLM, Multimodal)

Overview

• Do not run models
• Are used by execution tools to fetch weights and configs

What lives in an asset hub

• Model weights (LLMs, Stable Diffusion, audio, video)
• Adapters (LoRA, fine-tunes)
• Tokenizers and configs
• Example outputs and metadata
• Model cards, dataset cards, licenses, and tags
• Sometimes datasets
• Sometimes gated or approval-based access to specific assets

How asset hubs fit into the full stack

Asset hub → Execution tool → UI or API → Application

• Civitai → ComfyUI → Image generation
• Hugging Face → vLLM → LLM API
• Hugging Face → Ollama → Local chat

Glossary (Image generation assets)

• Simple: the main "brain" of the image generator. Without it, nothing works.
• Technical: a full set of trained diffusion model weights, now commonly distributed as .safetensors, with older ecosystems also using .ckpt, that defines the base image distribution, concepts, and capabilities.

• Simple: a small add-on that changes the model’s style or behavior without replacing the main brain.
• Technical: low-rank weight adapters applied on top of a base checkpoint at inference time to steer style, subjects, or concepts with minimal additional parameters.

• Simple: the part that turns the model’s internal representation into the final image you see.
• Technical: an encoder-decoder network that maps between pixel space and latent space, affecting color accuracy, contrast, and fine detail.

• Simple: a guide that forces the image to follow a pose, sketch, depth map, or structure.
• Technical: an auxiliary conditional network trained to inject structural constraints (pose, edges, depth, segmentation) into the diffusion process while keeping the base checkpoint unchanged.

Image-focused asset hubs

Civitai (Image)

• Community-driven hub for Stable Diffusion assets

• Stable Diffusion checkpoints
• LoRAs
• VAEs
• ControlNet models
• Example images and prompts

• Visual discovery
• Comparing styles via images
• Artist and creator workflows

• Execution tools: ComfyUI, Automatic1111, InvokeAI

• Finding a style-specific model
• Downloading LoRAs for a specific look
• Reusing community workflows

General-purpose asset hubs

Hugging Face Hub (LLM, Image, Multimodal)

• The primary open hub for machine learning models across all modalities
• At current scale, it spans over 2M models, 500k datasets, and 1M demo apps (Spaces)

• LLM weights and configs
• Stable Diffusion and diffusion models
• Audio and video models
• Tokenizers and datasets
• Model cards, dataset cards, metadata, and licenses
• Public, gated, and approval-based assets

• Engineers and researchers
• Programmatic access
• Versioning and reproducibility
• Git-based collaboration around models, datasets, and demos

• Execution tools: Ollama, llama.cpp, vLLM, TGI, ComfyUI
• Hugging Face-hosted inference options when users do not want to self-host everything

• Pulling official or upstream models
• Building production or research systems
• Accessing datasets alongside models
• Reviewing model cards before adoption
• Using hosted inference for faster evaluation or deployment

• Most Hugging Face hub content is Git-based, but the platform also supports storage buckets for assets that do not fit neatly into normal Git history

ModelScope (LLM, Image, Multimodal)

• A major general-purpose model hub and ecosystem platform
• Especially relevant for Chinese open-model ecosystems and related tooling

• LLMs and multimodal models
• Vision, audio, and video models
• Datasets, benchmarks, and related project assets

• Broad model discovery outside the Hugging Face-centric ecosystem
• Ecosystem integration around training, evaluation, and deployment
• Access to models that may appear there earlier or more prominently

• Its own ecosystem tooling, SDKs, and hosted services
• External inference and deployment workflows

• Finding models that are more visible in the ModelScope ecosystem
• Exploring Chinese-market or Alibaba-adjacent model releases
• Comparing hub availability across multiple ecosystems

Hybrid model catalogs and hosted inference

Replicate

• A hosted model platform with a public model catalog and API-first execution

• Fast experimentation
• Running community or official models without self-hosting
• Product integration through a managed API

• Testing image, video, or LLM workflows quickly
• Shipping features without managing inference infrastructure

fal

• A hosted inference platform with a broad catalog of production-ready generative models

• High-performance hosted inference
• Image, video, audio, and 3D generation workloads
• API-driven product use cases

• Building media-generation products on top of managed infrastructure
• Accessing fast hosted inference for popular generative models

OpenRouter

• A unified API and model-routing layer for many hosted LLM providers

• Comparing LLMs across providers
• Normalized access through one API
• Fast switching between commercial and open hosted models

• Evaluating LLMs side by side
• Building applications that need routing or fallback across providers

• OpenRouter is not a weight-distribution hub. It is an inference access layer.

Other common platforms

Tensor.Art

• An image-focused community platform closer to Civitai
• Most relevant for browsing creator-driven image models, LoRAs, and workflows

Kaggle Models

• Relevant inside the Kaggle ecosystem
• Better viewed as an ecosystem-specific model catalog than as a primary general-purpose hub

Rule of thumb

• If you care about styles and images: start with Civitai
• If you care about broad model discovery and engineering workflows: start with Hugging Face
• If you want another major general-purpose ecosystem, especially beyond the usual western-default tooling: check ModelScope
• If you want to run models quickly through an API, look at hybrid platforms like Replicate or fal
• If you want one API across many LLM providers, look at OpenRouter
• Execution usually happens elsewhere, but some platforms now bundle hub features with hosted inference

Comparison table