layout	page
title	🇺🇸 English
permalink	/
lang	en

SDXL Container

🇺🇸 English | 🇮🇳 हिंदी | 🇯🇵 日本語 | 🇨🇳 简体中文 | 🇪🇸 Español | 🇧🇷 Português (Brasil) | 🇰🇷 한국어 | 🇩🇪 Deutsch | 🇫🇷 Français

A docker container to train SDXL LoRA adapters and run SDXL inference.

This repo is optimized for “small image set” LoRA runs:

1. drop images into a folder
2. (optionally) auto-generate captions
3. train a LoRA into ./models/loras/
4. immediately generate images with that LoRA

---

What’s inside

• GPU trainer container
• Command entrypoint: train / caption / infer
• LoRA training wrapper
• Training launcher wrapper
• BLIP captioning tool
• Diffusers inference script
• CPU-only test container for CI

---

Architecture / Mounts

docker-compose.yml mounts local folders into the container:

• ./models → /models (base models + output LoRAs)
• ./datasets → /datasets (your raw images)
• ./workspace→ /workspace(runs + caches + outputs)
• ./scripts → /scripts (entrypoint + wrappers)

All commands run inside the container, but files are written to your host via these mounts.

---

Prerequisites

• Docker + Docker Compose
• GPU + toolkit (for gpus: all)
• An SDXL base model as either: (a) local .safetensors/diffusers dir under ./models/base/, or (b) a Hugging Face repo id (e.g., stabilityai/sdxl-turbo)
• A small dataset under ./datasets/<subject>/images/

---

Highlights:

• Reproducible: everything runs inside a container (no local Python env needed).
• Simple: one command to (optionally) caption images + train.
• Safe defaults for few-shot SDXL LoRA.
• Includes inference: SDXL txt2img with LoRA using diffusers.

---

Build

docker compose build trainer

---

Train (caption + LoRA)

# train
docker compose run --rm trainer train \
--base-model stabilityai/sdxl-turbo \
--images /datasets/yokosuka \
--run-name yokosuka \
--sdxl \
--caption-mode blip \
--concept-token sksyokosuka \
--max-train-steps 1600 \
--num-repeats 20 \
--network-dim 16 \
--network-alpha 8

Infer (txt2img)

docker compose run --rm trainer infer \
--base-model stabilityai/sdxl-turbo \
--lora /models/loras/title_***.safetensors \
--prompt "portrait photo of sksTitle, high detail, natural light" \
--negative-prompt "low quality, blurry, worst quality" \
--out-dir /workspace/outputs \
--num-images 4 \
--seed 123 \
--steps 30 \
--cfg 7.0 \
--lora-scale 0.8 \
--width 1024 --height 1024

Caption (BLIP)

If you want to generate .txt captions next to each image (same basename):

# caption
docker compose run  \
--rm trainer caption  \  
--images /datasets/title \   
--prefix sksSubject    \
--overwrite

Inference (SDXL txt2img with LoRA)

Generate images with the trained LoRA:

# inference
docker compose run  \
--rm trainer infer    \
--base-model /models/base/sd_xl_base_1.0.safetensors    \
--lora /models/loras/title_***.safetensors    \
--prompt "sksSubject seaside"    \
--negative-prompt ""    \
--out-dir /datasets/title/inference    \
--num-images 4    \
--steps 30    \
--cfg 7.0    \
--width 1024    \
--height 1024    \
--lora-scale 0.8    \
--seed 42

---

Test

docker compose -f docker-compose.test.yml build
docker compose -f docker-compose.test.yml run --rm test

LoRA algorithm

LoRA (Low-Rank Adaptation) fine-tunes a diffusion model by adding a low-rank update to selected weight matrices while keeping the base weights frozen.

For a weight matrix W, LoRA learns:

ΔW = (α / r) * (B @ A)

Where:

r is the rank (--network-dim)

α is the scaling factor (--network-alpha)

A and B are the low-rank trainable matrices

At inference time the effective weight becomes:

W' = W + ΔW

Additionally, this repo lets you control how strongly the LoRA influences generation via --lora-scale.

License

• Apache License 2.0