Qwen Image 2512 LoRA Trainer API Documentation

Qwen Image 2512 LoRA Trainer API Documentation

Playground

Qwen-Image-2512 LoRA Trainer lets you train custom LoRA models 10x faster with style, character, and object training. From concept to model in minutes, not hours—upload a ZIP file containing images to start. Ready-to-use REST inference API, best performance, no cold starts, affordable pricing.

Features

Qwen Image 2512 LoRA Trainer is a high-performance custom model training service for the Qwen Image 2512 text-to-image generation model. It allows you to train lightweight LoRA (Low-Rank Adaptation) adapters for personalized styles, characters, and concepts — with exceptional bilingual text rendering preserved throughout.

Training Architecture

The trainer is designed around Qwen Image’s 20B MMDiT architecture and produces specialized LoRA adapters optimized for the model’s unique capabilities:

• Base LoRA adapter Trains on the core Qwen Image representation to capture your target style, character, or object, while keeping the base model frozen and stable.

• Text-rendering preservation The training process is optimized to maintain Qwen Image’s superior Chinese and English text rendering capabilities even after fine-tuning.

• Bilingual prompt compatibility Trained LoRAs work seamlessly with both Chinese and English prompts, preserving the model’s multilingual strengths.

This architecture ensures that your LoRA:

• Remains compact and easy to share
• Is plug-and-play with supported UIs and pipelines (e.g. ComfyUI, AI Toolkit)
• Preserves the text rendering and bilingual capabilities of Qwen Image 2512

Training Process

1. Data Upload Prepare and upload a ZIP file containing your training images. Include 10-20 high-quality images for best results.

2. Configure Trigger Word Set a unique trigger word (e.g., “p3r5on”) that will activate your trained style or character in prompts.

3. Adjust Training Parameters

• steps — Total training iterations (default: 1000)
• learning_rate — Training speed (default: 0.0004)
• lora_rank — Adapter capacity (default: 16)

4. LoRA Training The system runs a tailored LoRA optimization loop:

• Freezes the base model weights
• Trains only the low-rank adapter layers
• Applies Qwen-optimized settings for best results

5. Model Export After training completes, you receive:

• A LoRA adapter file (.safetensors) compatible with Qwen Image 2512
• Ready to use with Qwen Image 2512 LoRA

Parameters

Pricing

Billing Rules

• Base price: $1 per 1,000 steps
• Total cost = $1 × (steps / 1000)
• Billed proportionally to the total number of steps in your job

Best Use Cases

• Character Consistency — Train on character images to maintain identity across generations.
• Brand Styles — Create custom visual styles for consistent marketing materials.
• Art Styles — Capture specific artistic aesthetics for creative projects.
• Product Visualization — Train on product photos for consistent e-commerce imagery.

Pro Tips

• Use 10-20 high-quality, diverse images of your subject for best results.
• Choose a unique trigger word that won’t conflict with common words.
• Start with default settings, then adjust if needed.
• Higher lora_rank captures more detail but increases file size.
• Lower learning_rate is more stable but requires more steps.

Notes

• Higher parameter values (steps, lora_rank) will increase training time.
• Training time scales with the number of images and total steps configured.
• For faster iterations, start with lower settings and increase gradually.

Try More Trainers

• Z-Image LoRA Trainer — High-performance LoRA trainer for Z-Image models with Turbo-compatible optimization.

• Wan 2.2 Image LoRA Trainer — LoRA trainer for the Wan 2.2 image model, ideal for custom styles that integrate into the Wan video/image ecosystem.

• Flux Dev LoRA Trainer — LoRA trainer tailored for the Flux Dev model, focusing on high-fidelity creative visuals.

Reference

• How to train your LoRA?
• How to use your LoRA?

Authentication

For authentication details, please refer to the Authentication Guide.

API Endpoints

Submit Task & Query Result

set -euo pipefail

export WAVESPEED_API_KEY="your-api-key"

REQUEST_BODY=$(cat <<'JSON'
{
  "data": "https://github.com/mdn/interactive-examples/archive/refs/heads/main.zip",
  "steps": 1000,
  "learning_rate": 0.0004,
  "lora_rank": 16
}
JSON
)

# 1. Submit the prediction.
SUBMIT_RESPONSE=$(curl --silent --show-error --fail-with-body \
  -X POST "https://api.wavespeed.ai/api/v3/wavespeed-ai/qwen-image-2512-lora-trainer" \
  -H "Authorization: Bearer ${WAVESPEED_API_KEY}" \
  -H "Content-Type: application/json" \
  -d "${REQUEST_BODY}")

TASK=$(printf '%s' "${SUBMIT_RESPONSE}" | jq 'if type == "object" and has("data") then .data else . end')
PREDICTION_ID=$(printf '%s' "${TASK}" | jq -r '.id // empty')
if [ -z "${PREDICTION_ID}" ]; then
  printf 'Submission response did not contain a prediction id
' >&2
  exit 1
fi
RESULT_URL=$(printf '%s' "${TASK}" | jq -r '.urls.get // empty')
if [ -z "${RESULT_URL}" ]; then RESULT_URL="https://api.wavespeed.ai/api/v3/predictions/${PREDICTION_ID}/result"; fi

# 2. Poll until the prediction finishes.
while true; do
  RESPONSE=$(curl --silent --show-error --fail-with-body \
    "${RESULT_URL}" \
    -H "Authorization: Bearer ${WAVESPEED_API_KEY}")
  RESULT=$(printf '%s' "${RESPONSE}" | jq 'if type == "object" and has("data") then .data else . end')
  STATUS=$(printf '%s' "${RESULT}" | jq -r '.status // empty')

  case "${STATUS}" in
    completed) printf '%s\n' "${RESULT}" | jq '.outputs'; break ;;
    failed|cancelled|timeout) printf '%s\n' "${RESULT}" | jq . >&2; exit 1 ;;
    created|processing) sleep 2 ;;
    *) printf 'Unexpected status: %s
' "${STATUS}" >&2; exit 1 ;;
  esac
done

Parameters

Task Submission Parameters

Request Parameters

Response Parameters

Result Request Parameters

Result Response Parameters