Muapi's AI LoRA trainer API fine-tunes a portable LoRA adapter from your own image dataset without GPU rental or training infrastructure. Upload a ZIP archive, choose a trigger word, and receive a downloadable .safetensors file ready for compatible image generation.Eight dedicated training endpoints cover FLUX.1 [dev], FLUX.2 Klein 4B/9B, Qwen-Image, Qwen-Image-2512, and Z-Image / Z-Image Base. Use one REST pattern with flat per-run or per-step pricing from $2.00 per 1,000 steps.
LoRA training is live on Muapi
Choose a base model, submit a ZIP dataset and trigger word, then poll the returned request ID for a downloadable LoRA file.
Train a custom Flux LoRA from 10–50 images and a trigger phrase. The flat $3.20 run fee covers the training job, and the resulting .safetensors file can be reused with compatible Flux Dev generation.
Train a style-focused adapter on FLUX.1 [dev] from a ZIP dataset, with downloadable safetensors output and instant platform AIR deployment.
Train style adapters on the smaller, faster FLUX.2 Klein 4B architecture with ZIP input, downloadable output, and instant AIR deployment.
Train higher-quality style adapters on the larger FLUX.2 Klein 9B architecture with per-step billing and instant AIR deployment.
Fine-tune Qwen-Image on a ZIP dataset for character, subject, or artistic-style consistency with configurable steps, learning rate, and rank.
Fine-tune the newer Qwen-Image-2512 checkpoint from a ZIP dataset at the same low per-step price as the base Qwen trainer.
Train custom Z-Image adapters from a ZIP archive with auto-tuned defaults and optional control over steps, learning rate, and LoRA rank.
Fine-tune the full non-distilled Z-Image Base checkpoint with auto-tuned defaults and manual hyperparameter control.
A LoRA (Low-Rank Adaptation) is a small, portable weights file that teaches a base image model a character, subject, brand identity, or visual style without retraining the full model. An AI LoRA training API turns that process into one HTTP request: upload a ZIP of reference images, submit the job, and poll for a downloadable .safetensors file.
Muapi exposes eight trainers through the same REST pattern used across the platform. Train once, reuse the resulting adapter across compatible generations on Muapi, or download the standard weights file for your own pipeline.
Bundle 4–50 training images and optional same-name .txt caption files into a ZIP at a publicly reachable URL.
Choose a short, unusual phrase that activates your trained subject or style in future prompts.
Train on FLUX.1 [dev], FLUX.2 Klein 4B/9B, Qwen-Image, Qwen-Image-2512, or Z-Image / Z-Image Base.
Use auto-tuned defaults or control steps, learning rate, and LoRA rank where the selected trainer supports them.
Download a standard .safetensors file that is not locked to Muapi's inference layer.
Trained adapters register for downstream Muapi inference without a separate publish step.
Teach a compatible image model a recurring person, mascot, character, or object for repeatable generations.
Capture a visual language, product treatment, or campaign aesthetic for future creative production.
Train on a curated illustration or photography set and reuse the learned look across new prompts.
Use low-cost Qwen or Z-Image trainers to test a dataset and trigger phrase before scaling up.
| Trainer | Base model | Price | Best For |
|---|---|---|---|
| Flux LoRA Trainer | FLUX.1 [dev] | $3.20/run | Subjects and styles with flat per-run pricing |
| FLUX.1 [dev] Style | FLUX.1 [dev] | $2.50/1,000 steps | Style-focused training with per-step billing |
| FLUX.2 Klein 4B Style | FLUX.2 Klein 4B | $4.50/1,000 steps | Faster iteration on the smaller architecture |
| FLUX.2 Klein 9B Style | FLUX.2 Klein 9B | $4.00/1,000 steps | Higher-quality style adapters |
| Qwen-Image | Qwen-Image | $2.00/1,000 steps | Low-cost subject or style training |
| Qwen-Image-2512 | Qwen-Image-2512 | $2.00/1,000 steps | The newer Qwen checkpoint at low cost |
| Z-Image | Z-Image | $2.50/1,000 steps | Fast iteration with auto-tuned defaults |
| Z-Image Base | Z-Image Base | $2.50/1,000 steps | Full non-distilled checkpoint fine-tuning |
Host a ZIP dataset, submit it to the trainer that matches your base model, save the request ID, and poll the standard result endpoint for lora_url.
curl -X POST https://api.muapi.ai/api/v1/flux-lora-trainer \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"images_data_url":"https://example.com/dataset.zip","trigger_phrase":"myxyz","training_style":"subject"}'
# Response: {"request_id":"REQUEST_ID"}The Flux trainer accepts images_data_url, a trigger_phrase, and an optional training_style of subject or style.
curl -X POST https://api.muapi.ai/api/v1/z-image-lora-trainer \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"data":"https://example.com/dataset.zip","trigger_word":"my_style","steps":2000,"learning_rate":0.0001,"lora_rank":16}'Z-Image, Qwen, and style trainers expose ZIP input plus step and hyperparameter controls appropriate to that family.
curl https://api.muapi.ai/api/v1/predictions/REQUEST_ID/result \ -H "x-api-key: YOUR_API_KEY" # Read the generated LoRA URL or URLs from the completed response and pass them to a compatible image model.
Poll until the task completes, then download lora_url or pass it to a compatible generation endpoint.
It trains a custom Low-Rank Adaptation model from your image dataset through a REST endpoint. Upload a ZIP, choose a trigger word, and receive a reusable .safetensors adapter.
Muapi currently exposes Flux.1 [dev], FLUX.2 Klein 4B and 9B, Qwen-Image, Qwen-Image-2512, and Z-Image / Z-Image Base across eight training endpoints.
Pricing starts at $2.00 per 1,000 steps for Qwen-Image and Qwen-Image-2512, is $2.50 per 1,000 steps for Z-Image and FLUX.1 [dev] Style, and reaches $4.50 per 1,000 steps for FLUX.2 Klein 4B. Flux LoRA Trainer costs a flat $3.20 per run.
Most trainers accept 4–50 images. Ten to thirty varied images are a practical starting point for a subject; keep the aesthetic consistent when training a style.
No. Caption .txt files are optional, but when included they should use the same base name as their image. Follow the trainer's trigger-word rules for placeholders.
You receive a downloadable .safetensors weights file and, for supported trainers, instant registration for Muapi inference.
Yes. The .safetensors output is a standard adapter for compatible pipelines using the matching base model architecture.
Create one Muapi API key, upload your dataset, and turn a visual concept into a reusable adapter.