fal-train

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

fal-train

Train custom LoRA models on fal.ai for personalized AI generation.

在fal.ai上训练自定义LoRA模型，实现个性化AI生成。

Scripts

脚本

Script	Purpose
`train.sh`	Submit a LoRA training job

脚本	用途
`train.sh`	提交LoRA训练任务

Usage

使用方法

Train Image LoRA (Style/Subject/Person)

训练图像LoRA（风格/主体/人物）

bash

./scripts/train.sh --images-url "https://example.com/training-images.zip" --trigger-word "sks style" --model fal-ai/flux-lora-fast-training

bash

./scripts/train.sh --images-url "https://example.com/training-images.zip" --trigger-word "sks style" --model fal-ai/flux-lora-fast-training

Train Portrait LoRA

训练肖像LoRA

bash

./scripts/train.sh --images-url "https://example.com/face-photos.zip" --trigger-word "ohwx person" --model fal-ai/flux-lora-portrait-trainer

bash

./scripts/train.sh --images-url "https://example.com/face-photos.zip" --trigger-word "ohwx person" --model fal-ai/flux-lora-portrait-trainer

Check Training Status

查看训练状态

bash

./scripts/train.sh --status --endpoint fal-ai/flux-lora-fast-training --request-id "abc123"

bash

./scripts/train.sh --status --endpoint fal-ai/flux-lora-fast-training --request-id "abc123"

Arguments

参数说明

Argument	Description	Required
`--images-url`	URL to zip of training images	Yes
`--trigger-word`	Word to activate the LoRA in prompts	Yes
`--model` / `-m`	Training model endpoint	No (default: fal-ai/flux-lora-fast-training)
`--steps`	Training steps	No (default: 1000)
`--status`	Check training job status	No
`--endpoint`	Endpoint for status check	With --status
`--request-id`	Request ID for status check	With --status
`--param`	Extra param as key=value (repeatable)	No

参数	说明	是否必填
`--images-url`	训练图像zip压缩包的URL	是
`--trigger-word`	提示词中激活LoRA的触发词	是
`--model` / `-m`	训练模型端点	否（默认值：fal-ai/flux-lora-fast-training）
`--steps`	训练步数	否（默认值：1000）
`--status`	查看训练任务状态	否
`--endpoint`	状态查询的端点	搭配--status使用
`--request-id`	状态查询的请求ID	搭配--status使用
`--param`	额外参数，格式为key=value（可重复传入）	否

Finding Models

查找模型

To discover the best and latest training/LoRA models, use the search API:

bash

undefined

如需查找优质的最新训练/LoRA模型，可使用搜索API：

bash

undefined

Search for LoRA training models

搜索LoRA训练模型

bash /mnt/skills/user/fal-generate/scripts/search-models.sh --query "lora training" bash /mnt/skills/user/fal-generate/scripts/search-models.sh --query "trainer" bash /mnt/skills/user/fal-generate/scripts/search-models.sh --query "fine-tune"


Or use the `search_models` MCP tool with keywords like "lora", "training", "trainer", "fine-tune".


或者搭配“lora”、“training”、“trainer”、“fine-tune”这类关键词使用`search_models` MCP工具。

Training Data Tips

训练数据提示

People: 10-20 photos, varied angles/lighting/expressions, consistent person
Styles: 10-15 images exemplifying the style, diverse subjects
Objects: 5-15 photos from different angles on various backgrounds
Images should be high quality, at least 512x512
Zip all images into a single .zip file and host at a URL

人物：10-20张照片，角度/光线/表情多样，确保为同一人
风格：10-15张能体现目标风格的图像，主体类型多样化
物体：5-15张在不同背景下从不同角度拍摄的目标物体照片
图像应为高质量，分辨率不低于512x512
将所有图像压缩为单个.zip文件并托管在可公开访问的URL上

Output Format

输出格式

json

{
  "diffusers_lora_file": {
    "url": "https://fal.media/files/.../lora.safetensors",
    "content_type": "application/octet-stream",
    "file_name": "lora.safetensors",
    "file_size": 12345678
  },
  "config_file": {
    "url": "https://fal.media/files/.../config.json"
  }
}

Use the

diffusers_lora_file.url

as the

lora_url

parameter when generating images with FLUX models.

json

{
  "diffusers_lora_file": {
    "url": "https://fal.media/files/.../lora.safetensors",
    "content_type": "application/octet-stream",
    "file_name": "lora.safetensors",
    "file_size": 12345678
  },
  "config_file": {
    "url": "https://fal.media/files/.../config.json"
  }
}

使用FLUX模型生成图像时，可将

diffusers_lora_file.url

作为

lora_url

参数传入。