Fine-Tuning an Image Generation Model with API

Overview

This guide walks you through fine-tuning an image generation model to create images in your custom style. You’ll learn how to:

Create an image generation fine-tune
Start the fine-tune run
Monitor training progress with sample outputs
Deploy the fine-tuned model
Run inference to generate images in your style

We’ll use one of the FLUX models, which are state-of-the-art for image generation:

base_model: black-forest-labs/FLUX.1-dev
script_type: image_generation

Your dataset should have two columns:

image_column – Training images showing your desired style
caption_column – Text descriptions of each image

For a quick minimal example, see the Image Generation Quick Start.

Prerequisites

Repository on Oxen with your training data committed, for example:
- Namespace: Tutorials
- Repository: CyberpunkArt
Dataset resource inside that repo, for example:
- main/train_images.parquet
- Each row contains an image path and descriptive caption
- Example: view sample dataset structure
API key with access to the repo:
- Exported as OXEN_API_KEY
Base URL for the Oxen API:
- Cloud: https://hub.oxen.ai
- Exported as OXEN_BASE_URL

Set these in your shell:

export OXEN_API_KEY="YOUR_API_KEY_HERE"
export OXEN_BASE_URL="https://hub.oxen.ai"
export OXEN_NAMESPACE="Tutorials"
export OXEN_REPO="CyberpunkArt"

Data Requirements

For best results with image generation fine-tuning:

Quantity: 10-50 images minimum, 100-500 images ideal
Quality: High resolution (1024x1024 or higher), consistent style
Captions: Descriptive prompts that explain what makes your images unique
Consistency: Images should share common elements (style, subject matter, theme)

Example dataset structure in train_images.parquet:

image	caption
images/001.jpg	a red sports car in cyberpunk style with neon lights
images/002.jpg	a cyberpunk city street at night with rain
images/003.jpg	a person wearing futuristic cyberpunk clothing

See the Parameter Guide to understand training duration and the Data Requirements section for detailed guidelines.

Step 1 – Create an Image Generation Fine-Tune

Endpoint

POST /api/repos/{owner}/{repo}/fine_tunes

For this example, we’ll use:

resource: main/train_images.parquet
base_model: black-forest-labs/FLUX.1-dev
script_type: image_generation

Training parameters:

image_column: image (your image column name)
caption_column: caption (your caption column name)
steps: 2000 (standard training duration)
learning_rate: 0.0002 (default for image models)
lora_rank: 16 (balanced capacity)
sample_every: 200 (generate samples every 200 steps to monitor progress)

Example curl request:

curl --location "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/repos/${OXEN_NAMESPACE:-Tutorials}/${OXEN_REPO:-CyberpunkArt}/fine_tunes" \
  -H "Authorization: Bearer ${OXEN_API_KEY}" \
  -H "Content-Type: application/json" \
  --data '{
    "resource": "main/train_images.parquet",
    "base_model": "black-forest-labs/FLUX.1-dev",
    "script_type": "image_generation",
    "training_params": {
      "image_column": "image",
      "caption_column": "caption",
      "steps": 2000,
      "batch_size": 1,
      "learning_rate": 0.0002,
      "lora_alpha": 16,
      "lora_rank": 16,
      "sample_every": 200,
      "samples": [
        {
          "prompt": "a sports car in cyberpunk style"
        },
        {
          "prompt": "a futuristic city street at night"
        }
      ],
      "timestep_type": "sigmoid",
      "use_lora": true
    }
  }'

The samples array allows you to specify test prompts that will be generated during training. This helps you monitor how well the model is learning your style.

The response will include a fine_tune object:

{
  "fine_tune": {
    "id": "ft_img_gen_12345",
    "status": "created",
    "resource": "main/train_images.parquet",
    "base_model": "black-forest-labs/FLUX.1-dev",
    "script_type": "image_generation",
    "training_params": { ... }
  }
}

Capture the fine-tune ID for the next steps:

FT_ID=$(curl --silent --location "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/repos/${OXEN_NAMESPACE:-Tutorials}/${OXEN_REPO:-CyberpunkArt}/fine_tunes" \
  -H "Authorization: Bearer ${OXEN_API_KEY}" \
  -H "Content-Type: application/json" \
  --data '{
    "resource": "main/train_images.parquet",
    "base_model": "black-forest-labs/FLUX.1-dev",
    "script_type": "image_generation",
    "training_params": {
      "image_column": "image",
      "caption_column": "caption",
      "steps": 2000,
      "batch_size": 1,
      "learning_rate": 0.0002,
      "lora_alpha": 16,
      "lora_rank": 16,
      "sample_every": 200,
      "samples": [
        {
          "prompt": "a sports car in cyberpunk style"
        },
        {
          "prompt": "a futuristic city street at night"
        }
      ],
      "timestep_type": "sigmoid",
      "use_lora": true
    }
  }' | jq -r '.fine_tune.id')

echo "Created fine-tune: $FT_ID"

Step 2 – Start the Fine-Tune Run

Once you have the fine_tune.id, trigger the training run. Endpoint

POST /api/repos/{owner}/{repo}/fine_tunes/{fine_tune_id}/actions/run

Example curl request:

curl --location "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/repos/${OXEN_NAMESPACE:-Tutorials}/${OXEN_REPO:-CyberpunkArt}/fine_tunes/${FT_ID}/actions/run" \
  -H "Authorization: Bearer ${OXEN_API_KEY}" \
  -X POST

The fine-tune will now begin training. This typically takes 1-2 hours for 2000 steps on a GPU.

For FLUX models, expect approximately 30-60 minutes per 1000 steps, depending on GPU availability and image complexity.

Step 3 – Monitor Fine-Tune Status and Sample Outputs

You can poll the fine-tune to check progress and view sample outputs generated during training. Endpoint

GET /api/repos/{owner}/{repo}/fine_tunes/{fine_tune_id}

Example monitoring script (bash):

while true; do
  RESP=$(curl --silent "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/repos/${OXEN_NAMESPACE:-Tutorials}/${OXEN_REPO:-CyberpunkArt}/fine_tunes/${FT_ID}" \
    -H "Authorization: Bearer ${OXEN_API_KEY}")

  echo "$RESP" | jq '.'

  STATUS=$(echo "$RESP" | jq -r '.fine_tune.status')
  CURRENT_STEP=$(echo "$RESP" | jq -r '.fine_tune.current_step // 0')

  echo "Status: $STATUS"
  echo "Current Step: $CURRENT_STEP / 2000"

  # Check for sample outputs (generated every 200 steps)
  SAMPLES=$(echo "$RESP" | jq -r '.fine_tune.sample_outputs // empty')
  if [ ! -z "$SAMPLES" ]; then
    echo "Sample outputs available:"
    echo "$SAMPLES" | jq -r '.[] | "  - \(.url)"'
  fi

  if [ "$STATUS" = "completed" ]; then
    OUTPUT_RESOURCE=$(echo "$RESP" | jq -r '.fine_tune.output_resource')
    echo "Fine-tune completed! Output: $OUTPUT_RESOURCE"
    break
  elif [ "$STATUS" = "errored" ]; then
    ERROR_MSG=$(echo "$RESP" | jq -r '.fine_tune.error')
    echo "Fine-tune failed: $ERROR_MSG"
    exit 1
  elif [ "$STATUS" = "stopped" ]; then
    echo "Fine-tune was stopped"
    break
  fi

  # Wait 30 seconds before checking again
  sleep 30
done

Understanding Training Progress

As training progresses, you’ll see:

Status updates: created → running → completed
Current step: Progress counter (e.g., 400/2000)
Sample outputs: Generated images at steps 200, 400, 600, etc.

Review the sample outputs to see how well the model is learning your style. The images should progressively match your training style better as training continues.

If sample outputs aren’t matching your style by step 1000, consider adjusting learning_rate or training for more steps. See the Parameter Guide for tuning advice.

Step 4 – Deploy the Fine-Tuned Model

Once training completes, deploy your model to a GPU-backed inference endpoint. Endpoint

POST /api/repos/{owner}/{repo}/fine_tunes/{fine_tune_id}/deploy

Example curl request:

DEPLOY_RESPONSE=$(curl --silent --location "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/repos/${OXEN_NAMESPACE:-Tutorials}/${OXEN_REPO:-CyberpunkArt}/fine_tunes/${FT_ID}/deploy" \
  -H "Authorization: Bearer ${OXEN_API_KEY}" \
  -X POST)

echo "$DEPLOY_RESPONSE" | jq '.'

The response will include deployment information with a model identifier you’ll use for inference:

{
  "deployment": {
    "model_slug": "oxen:tutorials/cyberpunkart-ft_img_gen_12345",
    "status": "deploying",
    "endpoint": "https://hub.oxen.ai/api/ai/images/generate"
  }
}

Capture the model slug:

DEPLOYED_MODEL=$(echo "$DEPLOY_RESPONSE" | jq -r '.deployment.model_slug')
echo "Deployed model: $DEPLOYED_MODEL"

Deployment may take 2-5 minutes as the model is loaded onto a GPU instance. You can check deployment status by polling the fine-tune endpoint.

Step 5 – Generate Images with Your Fine-Tuned Model

Now you can generate images in your custom style using the inference API. Endpoint

POST /api/ai/images/generate

Example curl request (text-to-image):

curl -X POST \
  "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/ai/images/generate" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${OXEN_API_KEY}" \
  -d "{
    \"model\": \"${DEPLOYED_MODEL}\",
    \"prompt\": \"a motorcycle racing through the city in cyberpunk style\",
    \"num_inference_steps\": 28,
    \"guidance_scale\": 7.5,
    \"width\": 1024,
    \"height\": 1024
  }"

Generate Multiple Images

You can generate multiple variations by setting num_images:

curl -X POST \
  "${OXEN_BASE_URL:-https://hub.oxen.ai}/api/ai/images/generate" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${OXEN_API_KEY}" \
  -d "{
    \"model\": \"${DEPLOYED_MODEL}\",
    \"prompt\": \"a futuristic building in cyberpunk style with neon signs\",
    \"num_inference_steps\": 28,
    \"guidance_scale\": 7.5,
    \"num_images\": 4,
    \"width\": 1024,
    \"height\": 1024
  }"

Inference Parameters

Parameter	Description	Typical Values
`prompt`	Text description of desired image	Any descriptive text
`num_inference_steps`	Quality vs speed (higher = better)	20-50 (28 is balanced)
`guidance_scale`	How closely to follow prompt	5-10 (7.5 is balanced)
`width` / `height`	Output resolution	512, 768, 1024
`num_images`	Number of variations to generate	1-4
`seed`	Random seed for reproducibility	Any integer

Use higher num_inference_steps (40-50) for final production images, and lower values (20-28) for quick iterations during testing.

Example Response

{
  "images": [
    {
      "url": "https://hub.oxen.ai/api/files/...",
      "width": 1024,
      "height": 1024
    }
  ],
  "parameters": {
    "model": "oxen:tutorials/cyberpunkart-ft_img_gen_12345",
    "prompt": "a motorcycle racing through the city in cyberpunk style",
    "num_inference_steps": 28,
    "guidance_scale": 7.5
  }
}

Complete Python Example

Here’s a complete Python script that ties everything together:

import requests
import time

BASE_URL = "https://hub.oxen.ai"
API_KEY = "YOUR_API_KEY"
NAMESPACE = "Tutorials"
REPO = "CyberpunkArt"

headers = {
    "Authorization": f"Bearer {API_KEY}",
    "Content-Type": "application/json"
}

# Step 1: Create fine-tune
print("Creating fine-tune...")
create_url = f"{BASE_URL}/api/repos/{NAMESPACE}/{REPO}/fine_tunes"
data = {
    "resource": "main/train_images.parquet",
    "base_model": "black-forest-labs/FLUX.1-dev",
    "script_type": "image_generation",
    "training_params": {
        "image_column": "image",
        "caption_column": "caption",
        "steps": 2000,
        "batch_size": 1,
        "learning_rate": 0.0002,
        "lora_rank": 16,
        "lora_alpha": 16,
        "sample_every": 200,
        "samples": [
            {"prompt": "a sports car in cyberpunk style"},
            {"prompt": "a futuristic city street at night"}
        ],
        "timestep_type": "sigmoid",
        "use_lora": True
    }
}

response = requests.post(create_url, headers=headers, json=data)
fine_tune_id = response.json()["fine_tune"]["id"]
print(f"Created fine-tune: {fine_tune_id}")

# Step 2: Start training
print("Starting training...")
run_url = f"{create_url}/{fine_tune_id}/actions/run"
requests.post(run_url, headers=headers)

# Step 3: Monitor progress
print("Monitoring progress...")
status_url = f"{create_url}/{fine_tune_id}"
while True:
    response = requests.get(status_url, headers=headers)
    fine_tune = response.json()["fine_tune"]
    status = fine_tune["status"]
    current_step = fine_tune.get("current_step", 0)

    print(f"Status: {status}, Step: {current_step}/2000")

    if status == "completed":
        print(f"Training completed! Output: {fine_tune['output_resource']}")
        break
    elif status == "errored":
        print(f"Training failed: {fine_tune.get('error')}")
        exit(1)

    time.sleep(30)

# Step 4: Deploy model
print("Deploying model...")
deploy_url = f"{status_url}/deploy"
response = requests.post(deploy_url, headers=headers)
deployed_model = response.json()["deployment"]["model_slug"]
print(f"Deployed: {deployed_model}")

# Wait for deployment
time.sleep(60)

# Step 5: Generate image
print("Generating image...")
generate_url = f"{BASE_URL}/api/ai/images/generate"
gen_data = {
    "model": deployed_model,
    "prompt": "a motorcycle racing through the city in cyberpunk style",
    "num_inference_steps": 28,
    "guidance_scale": 7.5,
    "width": 1024,
    "height": 1024
}

response = requests.post(generate_url, headers=headers, json=gen_data)
image_url = response.json()["images"][0]["url"]
print(f"Generated image: {image_url}")

Troubleshooting

Images not loading during training

Ensure image paths in your parquet file are relative to your repository root, or use full URLs. Verify that all images are committed to your Oxen repository with oxen status.

Out of memory errors

Reduce batch_size to 1 (default). If still failing, try reducing lora_rank to 8. See the Batch Size guide for more memory optimization tips.

Sample outputs don't match my style

Train for more steps (3000-5000 instead of 2000)
Ensure captions clearly describe the unique aspects of your style
Increase dataset size (100+ images recommended)
Try adjusting learning_rate (see Learning Rate guide)

Training taking too long

FLUX.1-dev takes ~1-2 hours for 2000 steps on GPU
Start with 1000 steps for quick testing
Consider using a faster model like Qwen/Qwen-Image for iteration
See supported models

Generated images have artifacts or low quality

Ensure training images are high resolution and consistent quality
Increase num_inference_steps to 40-50 during generation
Try different guidance_scale values (7.0-9.0)
Train for more steps to improve model quality

Next Steps

Advanced Parameters: See the Image Generation API Reference for all available parameters
Parameter Tuning: Learn about LoRA configuration and learning rate optimization
More Examples: Check out the Quick Start guide for simplified examples
Other Modalities: Explore Image Editing or Video Generation

With these skills, you can now fine-tune image generation models for any visual style, brand identity, or artistic direction!

Fine-Tuning API

API Reference

fine_tunes

evaluations

Fine-Tuning an Image Generation Model with API

Overview

Prerequisites

Data Requirements

Step 1 – Create an Image Generation Fine-Tune

Step 2 – Start the Fine-Tune Run

Step 3 – Monitor Fine-Tune Status and Sample Outputs

Understanding Training Progress

Step 4 – Deploy the Fine-Tuned Model

Step 5 – Generate Images with Your Fine-Tuned Model

Generate Multiple Images

Inference Parameters

Example Response

Complete Python Example

Troubleshooting

Next Steps

​Overview

​Prerequisites

​Data Requirements

​Step 1 – Create an Image Generation Fine-Tune

​Step 2 – Start the Fine-Tune Run

​Step 3 – Monitor Fine-Tune Status and Sample Outputs

​Understanding Training Progress

​Step 4 – Deploy the Fine-Tuned Model

​Step 5 – Generate Images with Your Fine-Tuned Model

​Generate Multiple Images

​Inference Parameters

​Example Response

​Complete Python Example

​Troubleshooting

​Next Steps

Overview

Prerequisites

Data Requirements

Step 1 – Create an Image Generation Fine-Tune

Step 2 – Start the Fine-Tune Run

Step 3 – Monitor Fine-Tune Status and Sample Outputs

Understanding Training Progress

Step 4 – Deploy the Fine-Tuned Model

Step 5 – Generate Images with Your Fine-Tuned Model

Generate Multiple Images

Inference Parameters

Example Response

Complete Python Example

Troubleshooting

Next Steps