Plain-English explanation
A base model is the common starting point beneath a family of specialized models. Pretraining gives it general statistical knowledge; later fine-tuning can teach conversational behavior, a domain, or an output style. A lightweight LoRA does not normally replace the base; it stores a smaller set of learned changes that only make sense when applied to a compatible model.
The term is contextual. In a model family discussion, it may identify the pretrained, non-chat release. In an image-generation download, “base model” may mean the checkpoint family and architecture an adapter was trained against.
From pretraining to a usable product
Pretraining optimizes model weights across broad training data. Developers can then instruction-tune or preference-tune a copy for chat, or train an adapter for a narrower behavior. The final interface may add a system prompt, retrieval, safety controls, and a chat template. Those layers matter, but they do not turn the interface into a new base model.
Why compatibility matters
Adapters and auxiliary components are usually trained against a particular architecture and parameter layout. Loading a LoRA onto the wrong image-model family can fail outright or produce poor output; the same concern applies to language-model adapters. Model cards and repository metadata should identify the required base, but labels such as “SD 1.5,” “SDXL,” or a particular LLM family are more useful than visual similarity alone.
Local platforms including ComfyUI and AUTOMATIC1111 expose this relationship when users combine checkpoints, VAEs, ControlNets, and LoRAs. A hosted product may hide the base entirely.
Base model, checkpoint, and foundation model
A checkpoint is a saved set of weights; it may be a base model or a later fine-tune. Foundation model emphasizes broad training and adaptability, while base model emphasizes the model's role as the starting dependency for something else. These labels overlap, so the repository's own documentation is more reliable than assuming every publisher uses them identically.
Limitations and misconceptions
- A derivative model does not inherit every capability or policy behavior unchanged.
- Matching a model name is not always enough; architecture, version, and sometimes text encoder must also match.
- “Base” does not mean neutral, unbiased, or unrestricted. It describes lineage, not a safety property.
Technical context
These SDXL-derived ecosystems show why ‘based on’ does not mean every adapter transfers across descendants.