A Stable Diffusion checkpoint is the main model file: one .safetensors (or older .ckpt) file that holds all the trained weights needed to make an image, including the denoising network, the text encoder and usually the VAE. Different checkpoints, such as SD 1.5, SDXL base or a fine-tune like Animagine XL, give different styles and need LoRAs and settings made for the same model family.
What a checkpoint is
In Stable Diffusion, “checkpoint” is the everyday name for a full model file. Hugging Face's Diffusers documentation calls this the single-file format: all the model weights stored in one .safetensors or .ckpt file, the layout ComfyUI and AUTOMATIC1111 expect. When you pick a model in those tools, you are picking a checkpoint.
ComfyUI's Load Checkpoint node shows what is inside. It returns three things: MODEL, the network that denoises the latent image; CLIP, the text encoder that turns your prompt into numbers; and VAE, which encodes and decodes images to and from the latent. Our guide to what a VAE is covers that last part.
Why it is called a checkpoint
The word comes from training: a checkpoint is the saved state of a model at a point in training, which can be picked up and trained further. The Stable Diffusion v1.5 model card shows this lineage. v1-5 was initialized from the v1-2 checkpoint and then fine-tuned for 595,000 steps at 512 × 512.
The same card also explains why one model can come as several files. v1-5-pruned-emaonly holds only the EMA weights, uses less VRAM and is meant for generating images; v1-5-pruned holds both EMA and non-EMA weights, uses more VRAM and is meant for further fine-tuning. For making pictures, the smaller inference file is the one you want.
Base models and fine-tuned checkpoints
A base model is the one its developer trained from a broad dataset, such as Stable Diffusion v1.5 or SDXL base 1.0. Many checkpoints people share are fine-tunes of a base model, trained further on a narrower set of images to push a style. Animagine XL 4.0, for example, says it was fine-tuned from SDXL 1.0 on about 8.4 million anime-style images at 1024 × 1024.
A fine-tune keeps its base model's architecture. That is what “model family” means in practice: an SDXL fine-tune is still an SDXL model, with SDXL's text encoders and expected image sizes. A checkpoint merge is a third route, blending several checkpoints into a new file; checkpoint merge vs LoRA explains how that works.
PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →
Checkpoint vs LoRA vs VAE
The checkpoint is the complete model and the only part you strictly need. A LoRA is a small add-on trained for one model family and loaded on top of a checkpoint at a weight you choose; see what a LoRA is. A separate VAE file replaces the decoder that is usually already inside the checkpoint.
Match the family across all three. A LoRA or VAE made for SD 1.5 is not meant for an SDXL checkpoint, and the reverse is also true. The file name and extension do not tell you the family: read the model page. What a .safetensors file is explains why the extension alone proves little.
.safetensors vs .ckpt
Both extensions can hold a checkpoint. Diffusers' documentation says older weights were commonly saved as .ckpt with Python's pickle, that pickled files may be unsafe because they can be exploited to run malicious code, and recommends safetensors instead. SDXL base 1.0, for example, is published as sd_xl_base_1.0.safetensors.
If a model is offered in both formats, take the .safetensors file. If it only comes as .ckpt, treat it like any other untrusted download.
How to load a checkpoint on desktop
In AUTOMATIC1111, put the file in models/Stable-diffusion (the repository includes a ‘Put Stable Diffusion checkpoints here’ note in that folder) and pick it from the checkpoint menu. In ComfyUI, put it in models/checkpoints and choose it in a Load Checkpoint node. In Diffusers, a single file loads with from_single_file on the matching pipeline, for example StableDiffusionXLPipeline for an SDXL checkpoint.
After switching checkpoints, check the model page for its recommended image size, sampler, steps and CFG, since a fine-tune can differ from its base model. Then rerun the same prompt and seed to compare.
PhoneDiffusion availability and article scope
PhoneDiffusion runs curated Core ML model packs on supported iPhone and iPad devices with iOS 17 or later. Each pack is a ready-to-run version of one checkpoint, mostly from the SD 1.5, SD 2.1, and SDXL families (for example Animagine XL 4), with its own VAE included. Many are faster few-step builds, so the app starts from the pack's own steps and guidance rather than the original model card's. Local generation is private and works offline once a pack is downloaded. Choosing a pack is choosing a checkpoint. There is no general production importer for arbitrary .safetensors or .ckpt checkpoints.
It also offers optional cloud models, such as FLUX and Seedream, on paid monthly or yearly plans; those are text-to-image only, and the text prompt is sent online to generate the image. The AUTOMATIC1111, ComfyUI and Diffusers steps above describe desktop and server tools.
Field notes
- A checkpoint is the full model in one file: denoising network, text encoder and usually the VAE.
- The name comes from training: a saved model state that can be trained further.
- Many shared checkpoints are fine-tunes of a base model such as SD 1.5 or SDXL, and keep its model family.
- LoRAs and VAEs must match the checkpoint's model family.
- Prefer .safetensors over pickle-based .ckpt files.
Try it on your phone
- Read the model page to find the checkpoint's base model and family.
- Download the .safetensors inference file (for SD 1.5, v1-5-pruned-emaonly, not v1-5-pruned) when there is a choice.
- Put it in models/Stable-diffusion (AUTOMATIC1111) or models/checkpoints (ComfyUI).
- Select it and apply the creator's recommended size, sampler, steps and CFG.
- Use only LoRAs and VAEs made for the same model family.
Sources and verification
Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.
- Hugging Face Diffusers: model formats (single file vs Diffusers format)A single-file checkpoint stores all model weights in one .safetensors or .ckpt file, for ComfyUI and AUTOMATIC1111; the Diffusers format splits UNet, text encoder, VAE and scheduler into folders; from_single_file loads one file; pickled .ckpt files may be unsafe.
- ComfyUI: Load Checkpoint node (nodes.py) and READMECheckpointLoaderSimple reads from models/checkpoints and returns MODEL (denoises latents), CLIP (encodes the prompt) and VAE (encodes and decodes images); the README says to put SD checkpoints in models/checkpoints.
- Stable Diffusion v1.5 model card: checkpoint history and filesv1-5 was initialized from the v1-2 checkpoint and fine-tuned for 595k steps at 512×512; v1-5-pruned-emaonly is EMA-only weights that use less VRAM for inference, v1-5-pruned holds EMA and non-EMA weights for fine-tuning.
- Cagliostro Research Lab: Animagine XL 4.0 model card (training)Fine-tuned from Stable Diffusion XL 1.0 on 8,401,464 anime-style images at 1024×1024.
- AUTOMATIC1111 web UI: models/Stable-diffusion folderRepository folder containing ‘Put Stable Diffusion checkpoints here.txt’; the README lists loading checkpoints in safetensors format and reloading checkpoints on the fly.
- Hugging Face: Safetensors documentationMaintainer documentation for the tensor-only file format, its layout, supported frameworks, and loading APIs.