Integrating ERNIE-Image LoRAs Into Your Brand System

If you ship marketing creative at real volume, your image model has to respect three constraints: your color palette, your typography feel, and your visual grammar (rounded vs sharp, flat vs gradient). A raw ERNIE-Image call gets you one on a good day. A brand-trained LoRA gets you all three on every call.

This walks the full loop: training on fal-ai/ernie-image-trainer, inference on fal-ai/ernie-image/lora and /lora/turbo, and real cost per asset.

Why a LoRA beats prompt engineering

You can get close with a long prompt. By render fifty, the rounded sans has turned gothic, the cream has shifted to beige, the grain has vanished. A LoRA bakes brand preferences into a small adapter. Thirty to a hundred training images teach ERNIE your palette, grammar, and typography bias. Once trained, it is a 50 to 200 MB file you plug into every inference call. The prompt becomes a scene description; brand feel comes from the weights.

Step 1: build the training set

Collect 30 to 100 brand-consistent images: posters, product shots, social graphics. Caption each focused on content, not style. 1024x1024 square is safest. Zip image-caption pairs and upload to fal storage.

Step 2: submit the training job

1import { fal } from "@fal-ai/client";
2
3fal.config({ credentials: process.env.FAL_KEY });
4
5const trainingData = await fal.storage.upload(
6  new File([await fetch("./brand-dataset.zip").then(r => r.blob())], "brand.zip")
7);
8
9const { request_id } = await fal.queue.submit("fal-ai/ernie-image-trainer", {
10  input: {
11    images_data_url: trainingData,
12    trigger_word: "brandstyle",
13    steps: 1500,
14    learning_rate: 0.0004,
15    resolution: 1024,
16    is_style: true
17  }
18});
19
20let status = await fal.queue.status("fal-ai/ernie-image-trainer", { requestId: request_id });
21while (status.status !== "COMPLETED") {
22  await new Promise(r => setTimeout(r, 30000));
23  status = await fal.queue.status("fal-ai/ernie-image-trainer", { requestId: request_id });
24}
25
26const result = await fal.queue.result("fal-ai/ernie-image-trainer", { requestId: request_id });
27const loraUrl = result.data.diffusers_lora_file.url;

The trigger_word activates the LoRA in every inference prompt. Pick a made-up word. is_style: true teaches visual style, not a subject. Training takes 15 to 40 minutes. 1500 steps on 60 images is the sweet spot for brand style.

Step 3: inference with the LoRA

1const result = await fal.subscribe("fal-ai/ernie-image/lora", {
2  input: {
3    prompt: "brandstyle launch poster, paper airplane over a city skyline, headline reads 'TAKEOFF', subhead reads 'April 2026'",
4    loras: [{ path: loraUrl, scale: 0.9 }],
5    image_size: "portrait_16_9",
6    num_inference_steps: 50,
7    guidance_scale: 4.5,
8    seed: 77
9  },
10  logs: true
11});

scale: 0.9 is LoRA strength. Drop to 0.6 if the LoRA overrides prompt content. For drafts, swap to /lora/turbo at 8 steps: one third the cost.

1await fal.subscribe("fal-ai/ernie-image/lora/turbo", {
2  input: {
3    prompt: "brandstyle social thumbnail, teal notebook on a cream desk",
4    loras: [{ path: loraUrl, scale: 0.9 }],
5    image_size: "landscape_16_9",
6    num_inference_steps: 8
7  }
8});

Step 4: cost math

Real numbers for a 50-asset campaign.

Training. 1500 steps on 60 images is $8 to $15. Call it $12.

Inference. /lora is $0.03/MP ($0.03 at 1024x1024, $0.055 at 1792x1024). /lora/turbo is $0.01/MP.

Fifty landscape assets with three variants each (150 renders) on standard: 150 * $0.055 = $8.25. Plus $12 training: $20.25 total.

Smarter path: 150 turbo drafts at $0.018 = $2.70, then promote 50 winners to standard at $2.75. Total $5.45 inference plus $12 training, roughly $17.50 all in, $0.35 per finished branded image. A freelance illustrator runs $200 to $500 per asset.

Step 5: the feedback loop

First LoRA ships; the teal is not quite your teal. Retrain with 10 to 20 new examples, bump to 2000 steps. Keep the trigger word. Swap the URL in production config; every render pulls updated weights.

One 60-image run captures palette, illustration grammar, and typography weight. Style lives in the LoRA; content lives in the prompt. Train once, infer a thousand times.