Outfit Studio

Note: This repository directory may be named clothless-next locally; the Python package and CLI are outfit-studio (outfit_studio/).

AI outfit inpainting for photos. Upload a person image, segment the person and clothing, optionally edit the masks, then inpaint a new outfit with Stable Diffusion.

The README demo assets are synthetic, fully clothed, and non-explicit. They show the vanilla repository path using runwayml/stable-diffusion-inpainting with ControlNet disabled.

What this repo does

• Serves a Gradio UI with login, credits, examples, history, and an ImageEditor mask workflow.
• Uses the FASHN Human Parser (SegFormer B4) to build person and clothing masks automatically.
• Lets users repaint masks before generation, so the model changes only intended clothing regions.
• Runs Diffusers inpainting with SD1.5, SDXL, or local .safetensors / .ckpt checkpoints.
• Optionally adds OpenPose ControlNet guidance for SD1.5 checkpoints.
• Saves pipeline debug artifacts for masks, crops, inpaint inputs, diffusion output, blending, metadata, and final results.

Vanilla demo

This walkthrough was produced from a fresh non-NSFW synthetic source image, then passed through the real pipeline with:

• model: runwayml/stable-diffusion-inpainting
• ControlNet: off
• seed: 4242
• steps: 24
• prompt target: a modest, fully clothed editorial outfit

The green overlay marks clothing pixels selected for inpainting; the red overlay marks the detected person area used for instance splitting and blending. Vanilla SD1.5 changes the outfit region cleanly, while custom checkpoints and ControlNet can improve style faithfulness for production use.

Useful demo files:

File	Purpose
docs/assets/demo-source.png	Synthetic, fully clothed input image
docs/assets/demo-mask-overlay.png	Person and clothing masks from segmentation
docs/assets/demo-inpaint-input.png	Crop prepared for inpainting
docs/assets/demo-diffusion-output.png	Raw vanilla SD1.5 inpaint result
docs/assets/demo-blended-crop.png	Feathered crop before full-frame compositing
docs/assets/demo-result-vanilla-sd15.png	Final feathered full-frame blend

Debug artifacts

Add these lines to .env when you want inspectable intermediates for a run:

OUTFIT_STUDIO_PIPELINE_DEBUG=true
OUTFIT_STUDIO_PIPELINE_DEBUG_DIR=debug-pipeline

Each run creates a timestamped folder. The segmentation phase records source and fused masks; the generation phase records the final masks, per-person crop, inpaint input, diffusion output, blended crop, full composite, and JSON metadata.

Requirements

• Python 3.10-3.12
• uv
• NVIDIA GPU recommended
• 8 GB+ VRAM recommended for SD1.5 inpainting

Quick start

cp .env.example .env
make install-fast
make run

Open http://localhost:7860.

Default login:

admin / admin

make install-fast creates the virtual environment from the pinned lockfile, installs dependencies (including CUDA PyTorch and ONNX Runtime GPU support), and downloads the configured segmentation and inpaint models.

Windows

powershell -ExecutionPolicy RemoteSigned -File scripts\setup.ps1
.\.venv\Scripts\Activate.ps1
.\.venv\Scripts\python.exe -m outfit_studio.main

Reinstall CUDA torch / ORT after uv sync: scripts\install-cuda.ps1. If the console mangles box-drawing glyphs, set OUTFIT_STUDIO_ASCII_LOG=true.

Make targets

Run make (or make help) for the colored target list. NO_COLOR=1 disables ANSI.

Target	Description
make / make help	Show targets
make install	Create .venv and install pinned Python dependencies
make install-fast	Install deps from uv.lock and download models
make download-models	Download segmentation weights and configured inpaint checkpoint
make run	Start the Gradio UI
make stop	Stop the local Gradio demo (make run)
make test	Run non-slow unit tests
make lint	Run Ruff
make docker-build	Build the Docker image
make docker-up	Start with GPU support
make docker-up-cpu	Start without GPU (base compose only)
make docker-download-models	Download ML weights into the running container
make docker-down	Stop containers
make docker-logs	Follow container logs

Docker

Requires Docker and, for GPU inference, the NVIDIA Container Toolkit.

cp .env.example .env
make docker-build
make docker-up          # GPU (compose + docker-compose.gpu.yml)
make docker-up-cpu      # CPU-only, no NVIDIA Container Toolkit
make docker-download-models   # if not bind-mounting local weights

Open http://localhost:7860. Default login (bootstrapped on first start):

admin / admin1234

Checkpoints bind-mount from ./models by default. Optional .env overrides:

• OUTFIT_STUDIO_HOST_MODELS_DIR — host checkpoint path (use the real path if ./models is a symlink)
• OUTFIT_STUDIO_HOST_HF_CACHE — Hugging Face hub cache (default: ~/.cache/huggingface)

Configuration

Runtime .env

Deployment settings live in .env: host, port, paths, auth, credits, logging, and pipeline debug flags.

Setting	Purpose
OUTFIT_STUDIO_HOST / OUTFIT_STUDIO_PORT	Server bind address
OUTFIT_STUDIO_MODELS_DIR	Local model/checkpoint directory
OUTFIT_STUDIO_OUTPUT_DIR	Generated image output directory
OUTFIT_STUDIO_DB_PATH	SQLite database path
OUTFIT_STUDIO_REQUIRE_AUTH	Login requirement
OUTFIT_STUDIO_DEFAULT_*	Default admin and credit settings
OUTFIT_STUDIO_DEBUG / OUTFIT_STUDIO_LOG_LEVEL	Diagnostics
OUTFIT_STUDIO_PIPELINE_DEBUG*	Intermediate artifact dumps
OUTFIT_STUDIO_TENSOR_TORRENT	Enable TT size-gate path (default true)
OUTFIT_STUDIO_TENSOR_TORRENT_MIN_PARAMS_GB	Skip TT below this param GiB (default 4.0)
OUTFIT_STUDIO_TENSOR_TORRENT_UNET	Allow UNet TT when size gate passes (default true)
OUTFIT_STUDIO_TENSOR_TORRENT_CACHE_DIR	TT artifact cache (default .cache/tensortorrent)
OUTFIT_STUDIO_CIVITAI_API_TOKEN	Bearer token for Civitai checkpoint downloads

Models, prompts, and generation defaults are intentionally not stored in .env.

Content and ML YAML

Branding, prompts, checkpoints, and generation tuning live in config/.

File	Tracked	Purpose
config/content.default.yaml	Yes	Vanilla defaults shipped with the repo
config/content.local.yaml	No	Local overrides for prompts, models, URLs, and generation tuning
config/content.local.yaml.example	Yes	Template for local overrides

cp config/content.local.yaml.example config/content.local.yaml

The local YAML file is merged on top of the default file. The shipped default inpaint checkpoint is runwayml/stable-diffusion-inpainting; the local example can point at custom .safetensors checkpoints and download URLs.

Stack

Component	Technology
UI	Gradio 4 + ImageSlider
Segmentation	FASHN Human Parser (SegFormer B4)
Pose, optional	rtmlib ONNX + ControlNet OpenPose
Inpainting	Diffusers SD1.5 / SDXL inpaint
Oversized models	TensorTorrent when params ≥ MIN_PARAMS_GB (else torch.compile)
Auth/history	SQLite + Argon2

When OUTFIT_STUDIO_TENSOR_TORRENT=true (default), Outfit Studio only hands a module to TensorTorrent if its parameters are at least OUTFIT_STUDIO_TENSOR_TORRENT_MIN_PARAMS_GB (default 4.0 GiB). SD1.5 inpaint UNets (~1.7 GiB bf16) and the human parser stay on eager / torch.compile — TT's compile + schedule tax made them slower, and a resident LATENCY plan OOMs next to ControlNet on 8 GB cards. Oversized UNets (e.g. SDXL ≈4.8 GiB) use TT: text_embeds / time_ids export as plain tensors; LATENCY first, MEMORY+NVMe streaming if capacity fails. ControlNet stays eager. Pose ONNX stays outside TT. Artifacts cache under .cache/tensortorrent/.

Development

make lint
make test
pre-commit install
pre-commit run --all-files

Python entry points:

.venv/bin/outfit-studio-download-models
.venv/bin/outfit-studio