No local GPU required
A Kohya SS alternative for people without a local GPU
Kohya SS is capable software and this is not a claim to be better than it. If you own a strong NVIDIA GPU and want control over every training parameter, schedulers included, install Kohya. NiceVois exists for the case where you do not have that hardware, or the install itself is the wall: a Python environment, CUDA versions, dependency pins, and a configuration screen with more than a hundred options standing between you and one model file.
Both approaches train an image LoRA and both produce a portable .safetensors file you can load in any compatible generator. The difference is where the computation happens and how much of the setup you own.
Where the two differ
| Kohya SS (local) | NiceVois (cloud) | |
|---|---|---|
| Hardware needed | NVIDIA GPU, typically 12–24 GB VRAM for modern bases | Any device with a browser |
| Setup | Python environment, CUDA, dependencies, base-model downloads | None |
| Configuration | Extensive: optimizers, schedulers, ranks, buckets | Images, a subject type, and a model name |
| Base model | Your choice: SD 1.5, SDXL, FLUX and more | Flex.1-alpha (Apache-2.0, FLUX-compatible) |
| Output | .safetensors | .safetensors in a ZIP with the training manifest and a usage guide |
| Captions | You caption and bucket the dataset yourself | Handled by the workflow |
| Runs offline | Yes | No |
| Batch and repeat runs | Unlimited once installed | A bounded experiment while we measure demand |
When Kohya SS is the better choice
- You already own the GPU and the environment works; the setup cost is behind you.
- You want a specific base model, a specific rank, or control over optimizers and schedulers.
- You train at volume, iterate on datasets daily, or need to work offline.
When one-click cloud training is the better choice
- You have the images and want the model, not a Python project.
- Your machine has no NVIDIA GPU, or not enough VRAM for a modern base.
- A Colab notebook already ate an afternoon with version errors.
- You want a result you can judge before deciding whether deeper tooling is worth learning.
What the experiment accepts
The current NiceVois image workflow takes 8–40 images in any common format — JPG, PNG, WEBP, HEIC, AVIF, BMP, TIFF, GIF — of at least 512 × 512, one subject type — person, product, character, or style — and a model name. Training runs on an isolated cloud GPU and returns a ZIP with the portable .safetensors LoRA, the exact training manifest, and a short usage guide with your trigger word. Files expire after seven days; the download is yours to keep.
Permission still applies
Train on images you own or have permission to use. That rule is the same whether the training happens in Kohya on your machine or here in the cloud — the tool changes, the responsibility does not.