Configuration

Latentry reads .env.local in its folder (the install script copies it from .env.example). The Settings page can set backends, gallery folders, the tagger and profile defaults too; what it saves (latentry.settings.json) takes precedence, and anything it leaves alone falls back to .env.local. Restart Latentry after editing .env.local.

Variable	Default	What it does
GEN_BACKENDS	—	Backends: id\|kind\|url\|profile, separated by ;
GEN_TOKEN_<ID>	—	A backend's token, by its id in upper case
GEN_MODEL_<ID>	first listed	A comfyui backend's checkpoint, as ComfyUI lists it
GEN_CONTROLNET_<ID>	—	A comfyui backend's Pose ControlNet, as ComfyUI lists it; unset: no pose
GALLERY_SAVE_DIR	./output	Where finished pictures are saved; unset: not saved
GALLERY_DIRS	—	Other folders to browse read-only, separated by ;
GALLERY_AUTO_TAG	off	1: tag each saved picture with WD14 as it arrives
WD14_MODEL_DIR	—	Folder of the built-in WD14 tagger (model.onnx, selected_tags.csv)
WD14_THRESHOLD, WD14_CHARACTER_THRESHOLD	0.35, 0.85	How sure a tag must be
TAGGER_BACKEND	automatic	builtin, or a backend id
A1111_TAGGER_MODEL	first "wd" model	The A1111 tagger extension's model
LATENTRY_HOST	127.0.0.1	0.0.0.0 to use Latentry from other devices
SETTINGS_EDIT	local	Who may change settings: local, lan or off
LATENTRY_SETTINGS_FILE	./latentry.settings.json	Where the Settings page saves. Model info (latentry.models.json) goes in the same folder
LATENTRY_ALLOW_UNC	off	1: let the Settings page use network folders (\\server\share)
ALLOWED_HOSTS	—	Extra hostnames /api answers for (behind a reverse proxy)
REQUEST_BUDGET_*	see below	Per-device request limits; REQUEST_BUDGET=off turns them off
LATENTRY_ENGINE	on	off: do not start the built-in engine with Latentry
LATENTRY_RUNTIME_DIR	./.runtime	Where the engine (uv, Python, PyTorch) is installed
HF_TOKEN	—	Only for models that need a Hugging Face login
PORT	3000	The install scripts' port

The engine also reads LATENTRY_DEVICE, LATENTRY_DTYPE, LATENTRY_OFFLOAD, LATENTRY_MAX_BATCH and LATENTRY_VRAM_FRACTION to override its GPU plan (see Setup and the engine), and LATENTRY_LORA_KEEP: how many LoRAs stay in the loaded model between runs, switched off, so calling one again is quick (default 4; 0 takes each out after its run).

The variables on this page, and the ones in .env.example and the guide, keep their names and meaning between releases; see Compatibility.

.env.example

Every variable, with its explanation:

# Copy to .env.local and fill in. Everything here is read by the server only;
# nothing reaches the browser except the backend ids.
#
# The Settings page (/settings) can set backends, gallery folders and profile
# defaults too. What it saves (latentry.settings.json) takes precedence; what
# it leaves alone falls back to the values below.

# ── Backends ────────────────────────────────────────────────────────────────
# One entry per backend, separated by ";":   id|kind|url|profile
#   id       what the UI calls it (a-z, 0-9, "-", "_")
#   kind     diffusers  a server speaking the API in docs/backend-api.md
#            a1111      AUTOMATIC1111 or Forge, started with --api
#            comfyui    a ComfyUI server (txt2img; it draws with the first
#                       checkpoint it lists, another is chosen on Settings)
#   url      where it listens
#   profile  generic | sdxl | illustrious | pony | anima  (the form's starting
#            point; it can be changed in the UI at any time)
GEN_BACKENDS=anima|diffusers|http://localhost:7865|anima;sdxl|a1111|http://localhost:7860|illustrious

# Tokens, by backend id in upper case (non-alphanumerics become "_").
#   diffusers: sent as "Authorization: Bearer <token>"
#   a1111:     "user:password", as given to --api-auth
#   comfyui:   ComfyUI has no login; a token is sent as a Bearer token, for a
#              proxy in front of it
# A token is only sent to the address GEN_BACKENDS gives its id; a backend
# moved elsewhere on the Settings page needs its token typed there.
# GEN_TOKEN_ANIMA=
# GEN_TOKEN_SDXL=user:password

# A comfyui backend's checkpoint, by backend id in upper case, as ComfyUI lists
# it (folders included). Unset: the first checkpoint ComfyUI lists.
# GEN_MODEL_COMFY=sdxl\base.safetensors

# ── WD14 tagging ────────────────────────────────────────────────────────────
# Latentry can run a WD14 tagger itself, on the CPU, with no backend involved.
# Download model.onnx and selected_tags.csv from a SmilingWolf WD tagger on
# Hugging Face (Apache-2.0), e.g. wd-eva02-large-tagger-v3 (best, ~1.2 GB) or
# wd-vit-tagger-v3 (smaller), into a folder of their own:
# WD14_MODEL_DIR=D:\models\wd-eva02-large-tagger-v3
# WD14_THRESHOLD=0.35
# WD14_CHARACTER_THRESHOLD=0.85
#
# What tags pictures: "builtin", a backend id, or unset for the built-in model
# when WD14_MODEL_DIR is set, else the first backend with a tagger.
# TAGGER_BACKEND=builtin
# A1111 / Forge with the WD14 tagger extension: which of its models to use.
# Default: the first one whose name contains "wd".
# A1111_TAGGER_MODEL=wd-v1-4-moat-tagger.v2

# Older single-backend form, still read when GEN_BACKENDS is unset:
# DIFFUSION_SERVER_URL=http://localhost:7865
# DIFFUSION_API_TOKEN=

# ── Gallery ─────────────────────────────────────────────────────────────────
# Where Latentry saves every finished image, with its settings embedded.
# Leave unset to not save (images can still be downloaded one by one).
GALLERY_SAVE_DIR=./output
# Other folders to browse, read-only, separated by ";" (e.g. ComfyUI's or the
# web UI's output folder). These must be on the machine running Latentry.
# GALLERY_DIRS=D:\ComfyUI\output;D:\stable-diffusion-webui\outputs\txt2img-images
# Tag each saved image with WD14 as it arrives (uses the tagger once per image).
# GALLERY_AUTO_TAG=1

# ── Settings page ───────────────────────────────────────────────────────────
# Who may change settings from the browser:
#   local (default)  only on this computer (http://localhost)
#   lan              from any address /api answers
#   off              nobody; edit this file or the settings file by hand
# SETTINGS_EDIT=local
# Where the settings page saves (it holds backend tokens; keep it private).
# LATENTRY_SETTINGS_FILE=./latentry.settings.json
# Network folders (\\server\share) on the settings page. Off: Windows signs in
# to a network folder's server as you when Latentry checks it, so whoever may
# change settings could choose where that sign-in goes. Folders set in this
# file are not affected.
# LATENTRY_ALLOW_UNC=1

# ── Access ──────────────────────────────────────────────────────────────────
# Latentry listens on this computer only (127.0.0.1). To use it from a phone
# or another computer on your network, listen on every interface, then open
# http://<this computer's address>:3000 there. On Windows, allow Node.js
# through the firewall on private networks when asked.
# LATENTRY_HOST=0.0.0.0
#
# Other devices must be paired before they can use it: Settings > Phones and
# other computers > Add a device shows a code to type on them (one use, 10
# minutes); a device stays paired for 30 days. Only turn pairing off behind a
# reverse proxy on this computer that does its own login.
# LATENTRY_PAIRING=off
#
# The key paired devices are recognised by, as 64 hex characters, instead of
# the file latentry.pairing-key. Set it at launch from your keychain (see the
# guide's "keep the pairing key in your keychain"); do not write it here.
# LATENTRY_PAIRING_KEY=
#
# /api answers only for localhost and private-network addresses. To serve it
# under a hostname (e.g. behind a reverse proxy), list the hostname here.
# ALLOWED_HOSTS=studio.home.arpa

# Per-IP request budget (per minute). REQUEST_BUDGET=off turns it off.
# REQUEST_BUDGET_PAGE_PER_MIN=120
# REQUEST_BUDGET_API_PER_MIN=300
# REQUEST_BUDGET_POST_PER_MIN=30
# REQUEST_BUDGET_GLOBAL_PER_MIN=1200
# REQUEST_BUDGET_SSE_PER_IP=12

# ── Latentry's own engine (see engine/README.md; set up from the Setup page) ──
# LATENTRY_ENGINE=off               # do not start the engine with Latentry
# LATENTRY_RUNTIME_DIR=./.runtime   # where uv, Python and PyTorch are installed
# LATENTRY_ENGINE_PACKAGES=latest   # install the newest Python packages, not the pinned ones
#                                   # (not hash-checked; for when a pinned one has a known
#                                   # vulnerability and no Latentry update has come yet)
# HF_TOKEN=hf_...                   # only for models that need a Hugging Face login