Qwen Image Text To Image LoRA API Documentation

Qwen Image Text To Image LoRA API Documentation

Playground

Qwen-Image LoRA is a 20B MMDiT next-gen text-to-image model with LoRA support for fast customization and refined image generation. Ready-to-use REST inference API, best performance, no coldstarts, affordable pricing.

Features

Qwen-Image-LoRA extends the base 20B MMDiT text-to-image model by allowing users to plug in custom LoRA weights (.safetensors) for fine-tuned control over style, characters, or artistic domains. This makes it a versatile tool for creators who want both world-class text rendering and personalized generation.

Why it looks great

• LoRA integration: Import external .safetensors LoRA weights and control blending strength via scale.
• SOTA text rendering: Rivals GPT-4o in English and is best-in-class for Chinese typography.
• In-pixel text generation: Text is seamlessly integrated into images (no overlays).
• Bilingual support: Handles Chinese & English with diverse fonts and complex layouts.
• General image excellence: Photorealistic, anime, impressionist, or minimalist styles—all supported.

Limits and Performance

• Max resolution per job: up to 1024 × 1024 pixels
• LoRA path: provide <owner>/<model-name> or external .safetensors URL
• LoRA scale: adjustable strength (default = 1.0)
• Output formats: JPEG / PNG / WEBP
• Processing speed: ~6–10 seconds per image
• Input prompt: supports multi-line descriptive text

Pricing

• $0.025 per image
• Each image is billed individually.

How to Use

1. Enter a prompt (supports detailed narrative & embedded text).
2. Set size (width & height, up to 1024×1024).
3. Add one or more LoRAs:

• Paste the path/URL of the LoRA .safetensors file.
• Adjust the scale (e.g., 0.5 for subtle effect, 1.0 for full strength).

4. (Optional) Set seed for reproducibility (-1 = random).
5. Choose output format (JPEG / PNG).
6. Run → preview results → iterate with different LoRA scales.

Pro tips for best quality

• Use specific LoRAs for characters, art styles, or IP consistency.
• Combine multiple LoRAs for hybrid results (e.g., anime + steampunk).
• Adjust scale carefully—too high may distort, too low may fade.
• Lock the seed to maintain subject consistency when swapping LoRAs.

Reference

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

Note

• Please use wavespeed-ai/qwen-image-lora-trainer to make sure your LoRA can use in this model!
• Or the corresponding model from official platform! (Civitai or Hugging Face)

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'
{
  "prompt": "A cinematic ocean wave at sunrise, highly detailed",
  "size": "1024*1024",
  "output_format": "jpeg"
}
JSON
)

# 1. Submit the prediction once. Never retry this request automatically:
# a retry after an unclear failure can create a second, billed task.
CHECK='The task may already exist. Check History or the predictions list before submitting again.'
if ! SUBMIT_RESPONSE=$(curl --silent --show-error --fail-with-body \
  --connect-timeout 10 --max-time 60 \
  -X POST "https://api.wavespeed.ai/api/v3/wavespeed-ai/qwen-image/text-to-image-lora" \
  -H "Authorization: Bearer ${WAVESPEED_API_KEY}" \
  -H "Content-Type: application/json" \
  -d "${REQUEST_BODY}"); then
  printf 'Submit failed. %s\n' "${CHECK}" >&2
  exit 1
fi

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 "${PREDICTION_ID}" ]; then
  printf 'Submission response did not contain a prediction id. %s\n' "${CHECK}" >&2
  exit 1
fi
RESULT_URL="https://api.wavespeed.ai/api/v3/predictions/${PREDICTION_ID}/result"

# 2. Poll until the prediction reaches a final status.
BODY_FILE=$(mktemp)
trap 'rm -f "${BODY_FILE}"' EXIT
while true; do
  # Transient errors (429, 5xx) are retried, honoring Retry-After.
  HTTP_CODE=$(curl --silent --show-error --fail-with-body \
    --connect-timeout 10 --max-time 60 --retry 5 --retry-delay 2 \
    --output "${BODY_FILE}" --write-out '%{http_code}' \
    "${RESULT_URL}" \
    -H "Authorization: Bearer ${WAVESPEED_API_KEY}" || true)
  case "${HTTP_CODE}" in
    2??) ;;
    000|429|5??)  # dropped connection or temporary error: keep polling
      printf 'Task %s: HTTP %s, polling again in 10s\n' "${PREDICTION_ID}" "${HTTP_CODE}" >&2
      sleep 10; continue ;;
    *) printf 'Task %s: HTTP %s\n' "${PREDICTION_ID}" "${HTTP_CODE}" >&2
      cat "${BODY_FILE}" >&2; exit 1 ;;
  esac
  RESULT=$(jq 'if type == "object" and has("data") then .data else . end' "${BODY_FILE}")
  STATUS=$(printf '%s' "${RESULT}" | jq -r '.status // empty')

  case "${STATUS}" in
    completed) printf '%s\n' "${RESULT}" | jq '.outputs'; break ;;
    failed|cancelled|timeout|deleted)
      printf 'Task %s ended with status %s\n' "${PREDICTION_ID}" "${STATUS}" >&2
      printf '%s\n' "${RESULT}" | jq . >&2; exit 1 ;;
    *) sleep 2 ;;
  esac
done

Parameters

Task Submission Parameters

Request Parameters

Every model also accepts the global request options enable_sync_mode and enable_base64_output; see their pages for details.

Response Parameters

Result Request Parameters

Result Response Parameters