A local, self-hosted pipeline for advanced image-to-image transformations
using Stable Diffusion (AUTOMATIC1111 WebUI) and Flux 1 Dev extensions.
Focused on realism evaluation, texture reconstruction, and anatomy-preserving edits via LoRA models.
This repository is a technical exploration of how diffusion models handle fine-grained image editing tasks, specifically:
- Reconstructing textures (e.g., skin, fabric) under prompt control
- Preserving facial identity and body pose consistency
- Tuning LoRA weight blends for style fidelity
By running everything locally, we avoid cloud dependencies and gain precise control over GPU memory, sampling, and prompt/seed configurations.
- Stable Diffusion WebUI by AUTOMATIC1111
- Flux 1 Dev extension
- Custom
.ckptand LoRA (.safetensors) model files - Python 3.10+, PyTorch, xformers, diffusers
⚙️ Tested on:
• Ubuntu 22.04 LTS
• NVIDIA RTX 3050 Ti (4 GB VRAM)
• 16 GB RAM (with 32 GB swap)
- ✅ Local GPU-powered inference (no cloud required)
- ✅ Image-to-image with adjustable strength, CFG scale, denoise, and sampler
- ✅ LoRA blending for controlled style transfers
- ✅ Prompt weighting and seed reproducibility
- ✅ VRAM optimization (
--medvram, attention slicing, CPU offload)
# 1. Clone this repo
git clone https://github.com/<your-username>/img2img-sd-flux-custom.git
cd img2img-sd-flux-custom
# 2. Drop custom models into ./models/
# e.g., models/flux1-dev.safetensors, models/my-lora.safetensors
# 3. Launch AUTOMATIC1111’s WebUI (inside a virtual environment):
cd ~/stable-diffusion-webui
python launch.py --xformers --medvram --opt-sub-quad-attention
# 4. In the WebUI’s “img2img” tab:
# • Upload ./images/input.jpg
# • Select your LoRA (e.g., “my-lora”) and adjust strength (0.3–0.7)
# • Tweak CFG (5–11), steps (50–100), sampler (Euler a, DDIM, etc.)
# • Click “Generate” → save to ./images/output.jpg
# 5. Compare side-by-side (assets/before-after.png)img2img-sd-flux-custom/
├── README.md
├── .gitignore
├── requirements.txt # (if using extra Python scripts)
├── models/
│ ├── flux1-dev.safetensors
│ └── my-lora.safetensors
├── images/
│ ├── input.jpg
│ └── output.jpg
├── config/
│ └── img2img-settings.json
├── scripts/
│ └── launch_notes.md
└── assets/
└── before-after.png
| Category | Details |
|---|---|
| 🧠 Prompt Engineering | Prompt weighting, seed control, sampler choice, CFG tuning |
| 🛠 VRAM Optimization | --medvram, attention slicing, CPU offload |
| 🎨 Style Control | LoRA blending (e.g., “my-lora.safetensors” for style fidelity) |
| 📊 Realism Benchmarking | Before/after comparisons, image fidelity metrics, texture checks |
This project uses open-domain models capable of sensitive content.
All testing is local and for technical evaluation only.
No explicit or disallowed content is shared. Intended for AI development and research.
Feel free to open an issue or submit a pull request.
Author: <Siddharthsinghkumar>
Repo: img2img-sd-flux-custom