SD3 FineTuning Setup Guide

## 🛠️ SD3 FineTuning Setup (Complete Guide)

### 1. ✅ Training Command

```bash
accelerate launch --mixed_precision bf16 --num_cpu_threads_per_process 1 sd3_train.py \
--save_model_as safetensors \
--train_data_dir "/workspace/sdxl/sd3/sd-scripts/exp_1_sd_3_training_Data" \
--in_json "/workspace/sdxl/sd3/sd-scripts/exp_1_sd_3_training_Data/meta_lat.json" \
--cache_latents_to_disk \
--cache_text_encoder_outputs_to_disk \
--pretrained_model_name_or_path /workspace/sdxl/sd3/sd3_weights/sd3.5_large.safetensors \
--clip_l /workspace/sdxl/sd3/sd3_weights/text_encoders/clip_l.safetensors \
--clip_g /workspace/sdxl/sd3/sd3_weights/text_encoders/clip_g.safetensors \
--t5xxl /workspace/sdxl/sd3/sd3_weights/text_encoders/t5xxl_fp16.safetensors \
--t5xxl_max_token_length 512 \
--fp8_base \
--highvram \
--seed 42 \
--gradient_checkpointing \
--mixed_precision bf16 \
--save_precision bf16 \
--output_dir /workspace/sdxl/sd3/exp_1_weights \
--output_name exp_1_sd3 \
--persistent_data_loader_workers \
--max_data_loader_n_workers 2 \
--optimizer_type adafactor \
--optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False" \
--lr_scheduler constant_with_warmup \
--max_grad_norm 0.0 \
--max_train_epochs 4 \
--save_every_n_steps 1000 \
--learning_rate 5e-5 \
--sdpa \
--fused_backward_pass \
--full_bf16
```

### 2. 📁 Dataset Directory Structure

```
/workspace/sdxl/sd3/sd-scripts/exp_1_sd_3_training_Data/
├── Image_1.JPG
├── Image_1.txt
├── Image_2.JPG
├── Image_2.txt
└── meta_lat.json
```

### 3. 🧾 JSON Metadata Format (`meta_lat.json`)

```json
{
"Image_1": {
"caption": "prompt text",
"train_resolution": [1216, 832]
},
"Image_2": {
"caption": "prompt text",
"train_resolution": [1216, 832]
}
}
```

Start with the referenced sd3_train.py entry point and compare its options with the supplied training command. Check the dataset layout and meta_lat.json format against the SD3 fine-tuning workflow; done means the setup guide is placed in the appropriate documentation location and accurately describes a usable command and metadata structure.

Written by the indexing model from the issue text.