Unlike older all-in-one SD 1.5/SDXL checkpoints that bundled text encoders, U-Net, and VAE into one file, modern DiT models like FLUX and open Video Generation Models (Sora, Veo, Kling, Wan) commonly ship modular checkpoints—separating the main diffusion weights from T5/CLIP text encoders and VAEs so users can mix FP8/GGUF precision independently in ComfyUI.
Checkpoints
Also known as: Model Checkpoints · Base Weights · Fine-tuned Checkpoints
Plain English
A saved snapshot of a neural network's learned numbers (weights) at a specific point in training. Downloading a 'checkpoint' gives you the complete, runnable brain of a model—whether an LLM or an image generator.
Technical Overview
During pre-training or fine-tuning, checkpoints serialize model parameter tensors (and optionally optimizer states for resuming training). In the open image/video ecosystem (Civitai, ComfyUI), 'checkpoint' specifically refers to a full standalone base or fine-tuned weight file—distinct from lightweight adapters like LoRA & QLoRA—typically distributed as Safetensors or quantized GGUF files on Hugging Face.
Chronological Entries (3)
Legacy .ckpt and .pt checkpoints relied on Python's pickle, which could execute arbitrary malicious code on load. Platforms like Hugging Face and Civitai migrated the ecosystem to Safetensors as the default checkpoint container, alongside GGUF for local CPU/consumer-GPU LLM inference.
In machine learning, a checkpoint is a serialized snapshot of model weights saved during or at the end of training. While pre-training clusters save checkpoints every few thousand steps for fault tolerance and evaluation, end users download release checkpoints to run inference or build downstream fine-tunes.
Direct Relations (5)
- UsesSafetensors
- UsesGGUF
- Alternative toLoRA & QLoRA
- Part ofHugging Face
- Part ofCivitai