Train a custom Flux LoRA from your own images. Upload a .zip of 10-50 photos (and optional captions) with a trigger phrase, and receive a portable .safetensors file you can plug into any Flux Dev pipeline for consistent subjects, characters, or styles.
About this model
Train a custom Flux LoRA from your own image collection in minutes. Upload a .zip of 10-50 photos and pick a unique trigger phrase; you'll receive a portable .safetensors file you can plug into any Flux Dev image-generation pipeline to reproduce a specific subject, character, or visual style with high fidelity.
Cost analysis
| Provider | Cost | Notes |
|---|---|---|
| muapiapp | $3.20 per training run | Flat fee per training run, regardless of zip size within the supported range. The trained LoRA is yours to reuse across unlimited generations. |
| Fal.ai | $2.00 per training run | Available as a standalone training endpoint but requires you to manage hosting, key rotation, and integration separately. |
| Replicate | Varies | Comparable Flux LoRA trainers exist but pricing is usage-based (per training second) and varies by hardware tier. |
Flat fee per training run, regardless of zip size within the supported range. The trained LoRA is yours to reuse across unlimited generations.
Available as a standalone training endpoint but requires you to manage hosting, key rotation, and integration separately.
Comparable Flux LoRA trainers exist but pricing is usage-based (per training second) and varies by hardware tier.
** Competitor pricing is estimated based on similar model architectures and usage tiers.
Configuration schema
| Parameter | Type | Description | Default |
|---|---|---|---|
| Images Zip URL | string | Public URL to a .zip archive containing 10-50 training images. Optional caption .txt files (same name as the matching image) can include a [trigger] placeholder which is replaced with trigger_phrase. | https://d3adwkbyhxyrtq.cloudfront.net/loradata/train_lora.zip |
| Trigger Phrase | string | A unique word or short phrase that activates the trained concept at inference time. Keep it short and unusual so it does not collide with normal vocabulary. | mysubject |
| Training Style | Enum (2 options) | What the LoRA learns: 'subject' (a person, character, or object) or 'style' (an aesthetic or art direction). | subject |
Public URL to a .zip archive containing 10-50 training images. Optional caption .txt files (same name as the matching image) can include a [trigger] placeholder which is replaced with trigger_phrase.
https://d3adwkbyhxyrtq.cloudfront.net/loradata/train_lora.zipA unique word or short phrase that activates the trained concept at inference time. Keep it short and unusual so it does not collide with normal vocabulary.
mysubjectWhat the LoRA learns: 'subject' (a person, character, or object) or 'style' (an aesthetic or art direction).
subjectDeveloper documentation
Prepare your images: Collect 10-50 high-quality images of the subject or style. For subjects, vary angles, expressions, and lighting; for styles, vary content but keep the aesthetic consistent.
Optionally add captions: For each image you can include a .txt file with the same base name containing a caption. Use the placeholder [trigger] anywhere you want the trigger phrase substituted in.
Zip and host the archive: Bundle the images (and optional captions) into a single .zip and upload it to a publicly reachable URL.
Pick a trigger phrase: Choose a short, unusual word (e.g. myxyz) that does not collide with normal English vocabulary. This phrase activates the LoRA at inference time.
Submit and wait: Training typically completes in 5-10 minutes. The webhook (or status poll) returns a lora_url pointing to the trained .safetensors file.
Use the LoRA: Pass lora_url to any Flux Dev LoRA-capable image-generation endpoint and include your trigger phrase in the prompt.
Frequently asked
At least 10 — more is usually better, up to about 50. For a subject, prioritise variety in pose and lighting. For a style, prioritise consistency in aesthetic.
No, captions are optional. If you do include them, name each `.txt` file identically to its image and use `[trigger]` wherever you want the trigger phrase injected.
Typically 5-10 minutes from submission to a downloadable `.safetensors` file. The endpoint is asynchronous — provide a `webhook_url` or poll the status endpoint.
It is the unique token you mention in inference prompts to activate the trained concept. Pick something short and unusual so it does not collide with normal vocabulary in your prompts.
Any image-generation endpoint that accepts a Flux Dev LoRA URL — including muapiapp's flux-dev-lora — can consume the resulting `.safetensors` file directly.