fal.ai ERNIE-Image LoRA Cloud Training API: Zero-GPU Fine-Tuning at $1.2 per Run

Summary: fal.ai has launched the official ERNIE-Image LoRA cloud training API (fal-ai/ernie-image-trainer) at $1.2 per 1,000 training steps. No local GPU required, no environment setup — upload a ZIP archive and train your custom LoRA. This article covers the complete API usage, dataset preparation, parameter tuning, output file handling, and a thorough comparison with local training, enabling you to fine-tune ERNIE-Image styles and characters without owning a GPU.

1. Why Cloud LoRA Training?

ERNIE-Image's LoRA fine-tuning capability has been extensively validated by the community (EI-071, EI-055), but local training faces three major barriers:

Hardware requirement: 8B-parameter DiT model LoRA training needs 24GB+ VRAM (RTX 3090/4090 or A100)
Environment setup: Aligning Diffusers, Accelerate, PEFT library versions is complex
Training time: 2,000 steps typically take 30-60 minutes locally

fal.ai's cloud API solves all three: no GPU needed, no setup, results in minutes, with commercial use allowed.

2. fal.ai ERNIE-Image Trainer API Overview

Endpoint: fal-ai/ernie-image-trainer

Feature	Details
Base Model	ERNIE-Image 8B (DiT)
Training Type	LoRA (Diffusers format)
Pricing	$1.2 / 1,000 steps
License	✅ Apache 2.0 / Commercial use OK
Recommended Steps	1,000-3,000
Output Format	`.safetensors` LoRA + config file

Cost Breakdown

Training Steps	Cost	Use Case
1,000 steps	$1.20	Style exploration / quick testing
2,000 steps	$2.40	Standard character/style LoRA (recommended)
3,000 steps	$3.60	High-quality / complex concept LoRA

3. Dataset Preparation: ZIP Archive Format

Directory Structure

training_data.zip
├── 001.jpg
├── 001.txt    ← "a portrait photo of [trigger word] wearing a blue shirt"
├── 002.jpg
├── 002.txt    ← "a portrait photo of [trigger word] smiling"
├── 003.jpg
├── 003.txt    ← "a portrait photo of [trigger word] in a suit"
└── ...

Key Rules

Images: Any common format (JPG/PNG), recommended 1024×1024 or higher
Captions: .txt file with the same name as the image
Trigger word: Use a unique identifier (e.g., spfg person) to avoid conflicts with common words
Image count: Recommend 15-30 images, covering different angles, lighting, and backgrounds

Caption Writing Tips

# ✅ Recommended: descriptive + trigger word "a professional portrait of spfg person, studio lighting, sharp focus, 8K detail" ❌ Avoid: too brief "person" ✅ Character training: include identity/style description

"a digital painting of spfg warrior, fantasy armor, dynamic pose, dramatic lighting"

Tip: If some images lack .txt files, set the default_caption parameter as a fallback.

4. Complete API Usage Workflow

4.1 Install Client

npm install --save @fal-ai/client

4.2 Configure API Key

export FAL_KEY="your_fal_api_key_here"

4.3 JavaScript Training Example

import { fal } from "@fal-ai/client";
// 1. Upload training data ZIP

const zipFile = new File([/* your ZIP data */], "training.zip", {

type: "application/zip",

});

const zipUrl = await fal.storage.upload(zipFile);
// 2. Submit training task (using queue protocol)

const { request_id } = await fal.queue.submit(

"fal-ai/ernie-image-trainer",

{

input: {

images_data_url: zipUrl,

steps: 2000,              // 2000 steps = $2.40

learning_rate: 0.0005,    // default learning rate

default_caption: null,    // no default caption

},

webhookUrl: "https://your-server.com/training-complete&quot;, // optional: callback

}

);
console.log("Training submitted, Request ID:", request_id);
// 3. Poll status (if not using webhook)

const checkStatus = async () => {

const status = await fal.queue.status("fal-ai/ernie-image-trainer", {

requestId: request_id,

logs: true,

});

console.log("Current status:", status.status);

return status.status === "COMPLETED";

};

4.4 Python Training Example

import fal_client
Upload ZIP file
with open("training_data.zip", "rb") as f:

zip_url = fal_client.upload_file(f)
Submit training task
request = fal_client.queue.submit(

"fal-ai/ernie-image-trainer",

arguments={

"images_data_url": zip_url,

"steps": 2000,

"learning_rate": 0.0005,

}

)
print(f"Training submitted: {request['request_id']}")
Wait for completion
result = fal_client.queue.result(

"fal-ai/ernie-image-trainer",

request_id=request["request_id"]

)
print("LoRA file:", result["diffusers_lora_file"]["url"])

print("Config file:", result["config_file"]["url"])

5. Training Parameters Deep Dive

5.1 Learning Rate

Learning Rate	Use Case	Risk
0.0001	Conservative training / large concepts	May underfit
0.0005	Default / general purpose	Balanced choice
0.001	Fast convergence / strong style	May overfit
0.002+	Special cases	High overfit risk

5.2 Training Steps

Steps	Cost	Effect
500-1,000	$0.60-$1.20	Initial learning, weak concept
1,500-2,000	$1.80-$2.40	Standard LoRA, recommended
2,500-3,000	$3.00-$3.60	Deep memory, watch for overfit
3,000+	$3.60+	Usually unnecessary

Rule of thumb: Start with 2,000 steps + 0.0005 learning rate, then adjust based on output quality.

6. Post-Training: Using the LoRA Files

6.1 Download Output Files

After training, the API returns two files:

diffusers_lora_file: LoRA weights (.safetensors format)
config_file: Model configuration

6.2 Using in Diffusers

from diffusers import ErnieImagePipeline
pipeline = ErnieImagePipeline.from_pretrained("baidu/ERNIE-Image")
Load LoRA
pipeline.load_lora_weights(

"path/to/diffusers_lora_file.safetensors",

adapter_name="my_custom_style"

)
Generate image
image = pipeline(

prompt="a spfg portrait, studio lighting",

lora_scale=0.8,  # LoRA strength

num_inference_steps=50,

).images[0]

image.save("output.jpg")

6.3 Using in ComfyUI

Download diffusers_lora_file to ComfyUI/models/loras/
Add a "Load LoRA" node in ComfyUI
Select the LoRA file, set strength (0.6-1.0)
Connect to the ERNIE-Image checkpoint node

7. fal.ai vs Local Training: Complete Comparison

Dimension	fal.ai Cloud API	Local Training
GPU Required	❌ None	✅ 24GB+ VRAM
Initial Cost	$0 (pay-per-use)	$1,000+ (GPU)
Per Training	$1.20-$3.60	~$0.10 electricity
Training Time	3-10 minutes	30-60 minutes
Setup	Zero config	Complex (Diffusers/PEFT/Accelerate)
Parameter Flexibility	Limited (4 params)	Full control
Commercial License	✅	✅
Best For	Beginners / occasional training	Pro users / high-frequency

When to choose fal.ai?

No GPU or insufficient VRAM
Quick LoRA concept testing
Occasional training (< 10 runs/month)
Don't want to handle environment setup

When to choose local training?

High-frequency training (> 10 runs/month, lower cost)
Need full control over training parameters
High dataset privacy requirements
Already have GPU resources

8. Practical Case: Training a Character Consistency LoRA

Background

Train a LoRA for a specific character, used for comic character consistency.

Step 1: Prepare Dataset

Collect 20 character images (different expressions/angles/clothing)
Caption each: "a comic illustration of spfg character, [scene description]"
Package as ZIP

Step 2: Submit Training

const { request_id } = await fal.queue.submit(
  "fal-ai/ernie-image-trainer",
  {
    input: {
      images_data_url: zipUrl,
      steps: 2000,
      learning_rate: 0.0005,
    },
  }
);

Step 3: Validate Results

Generate images with different prompts, check character consistency
If character features aren't prominent enough → increase steps or adjust learning rate
If overfitted (background/clothing memorized) → reduce steps

9. FAQ

Q: How long does training take?

A: 2,000 steps typically takes 3-10 minutes, depending on fal.ai server load.

Q: Can I train multi-concept LoRAs?

A: Yes, but use different trigger words to distinguish concepts.

Q: Can I use the output LoRA on other platforms?

A: Output is standard Diffusers LoRA format, compatible with ComfyUI, Diffusers API, etc.

Q: What if training fails?

A: Check ZIP format, ensure images have valid captions. fal.ai returns error logs.

Q: How does this differ from Z-Image Turbo Trainer?

A: ERNIE-Image Trainer is based on Baidu's 8B DiT ($1.2/1K steps), Z-Image Turbo Trainer on Alibaba's 6B ($0.85/1K steps). ERNIE-Image excels in text rendering and instruction following.

10. Conclusion

fal.ai's ERNIE-Image LoRA cloud training API opens the door for zero-GPU users to fine-tune ERNIE-Image. Starting at $1.2, training in minutes, zero config — style and character LoRA training has never been more accessible.

For occasional trainers, this is the most cost-effective choice; for high-frequency users, combine with local training (EI-071) and cloud training based on your needs.

Next step: Try your first LoRA training on fal.ai/ernie-image-trainer!

This article is based on fal.ai official API documentation and the ERNIE-Image technical report. API pricing and parameters may change — refer to the fal.ai official docs for the latest.

fal.ai ERNIE-Image LoRA Cloud Training API: Zero-GPU Fine-Tuning at $1.2 per Run

Table of Contents

fal.ai ERNIE-Image LoRA Cloud Training API: Zero-GPU Fine-Tuning at $1.2 per Run

1. Why Cloud LoRA Training?

2. fal.ai ERNIE-Image Trainer API Overview

Cost Breakdown

3. Dataset Preparation: ZIP Archive Format

Directory Structure

Key Rules

Caption Writing Tips

❌ Avoid: too brief

✅ Character training: include identity/style description

4. Complete API Usage Workflow

4.1 Install Client

4.2 Configure API Key

4.3 JavaScript Training Example

4.4 Python Training Example

Upload ZIP file

Submit training task

Wait for completion

5. Training Parameters Deep Dive

5.1 Learning Rate

5.2 Training Steps

6. Post-Training: Using the LoRA Files

6.1 Download Output Files

6.2 Using in Diffusers

Load LoRA

Generate image

6.3 Using in ComfyUI

7. fal.ai vs Local Training: Complete Comparison

When to choose fal.ai?

When to choose local training?

8. Practical Case: Training a Character Consistency LoRA

Background

Step 1: Prepare Dataset

Step 2: Submit Training

Step 3: Validate Results

9. FAQ

Q: How long does training take?

Q: Can I train multi-concept LoRAs?

Q: Can I use the output LoRA on other platforms?

Q: What if training fails?

Q: How does this differ from Z-Image Turbo Trainer?

10. Conclusion