Teach AI your face, your product, or your style
Upload 8–40 photos, say what they show, and press Train. NiceVois builds an image model of your subject, generates new pictures with it right here, and hands you the trained model file — yours to download and keep.
- Nothing to install — it all runs in the browser
- One press to start; ready in about an hour
- You keep the model file, and it works outside NiceVois
Your first model is free. One full-strength run on the house, at any image count and any training depth. Further runs are quoted before they start and draw on prepaid NiceVois balance, which is not currently sold — so the free run is the whole offer for a new account today.
Image model trainer
Build your model
One free experiment run
Sign in to train your image LoRA
Signing in keeps the result with one account and prevents the test budget from being abused.
Your click is enough. It helps us decide whether this should become the second NiceVois tool.
Keep this page open while the upload completes. Training continues in the cloud afterward.
Calculating…
Portable artifact ready
Your image LoRA is ready
The ZIP contains the trained .safetensors file, the exact training manifest, and a short usage guide.
Generate with your model
Describe what you want. NiceVois turns it into the language your model understands.
Renders are private to your account. Sexual content, content involving minors, graphic violence, and hate imagery are never rendered. Only depict people you have permission to depict.
Ten setup steps, reduced to three
Images in, portable model out
- 1Add 8–40 images
Use varied, clear images you own or have permission to use — any common format works.
- 2Name the subject
Choose person, product, character, or style. The training configuration stays out of your way.
- 3Train and keep the file
Download the LoRA package and a short guide. You own the artifact, not an account-locked preset.
For the technically minded
Under the hood: a portable FLUX-compatible LoRA
- What trains. A rank-16 LoRA on
Flex.1-alpha(Apache-2.0), driven by ai-toolkit, at 250–2000 steps with a per-subject recommendation — a face is budgeted like an identity, not like a style. - What you download. The
.safetensorsadapter itself with its trigger word, the exact training manifest, and preview samples. Load it in ComfyUI, Forge, or any FLUX-compatible runtime — it is not locked to NiceVois. - What the pipeline does. Your images are captioned with the trigger word, normalized to upright PNG or JPEG, trained on an isolated cloud GPU, and the finished package is size- and SHA-256-verified before delivery.
- Renders. On-page generation loads your adapter onto the same base model, and prompts are composed into the caption language the adapter was trained on.
What NiceVois is automating
Dataset validation, captions, dependency setup, model configuration, GPU execution, checkpoint selection, integrity checks, and output packaging.
What stays yours
Your source images remain private during the benchmark and expire after seven days. The finished LoRA is a file you can keep and use elsewhere.