Skip to content

Changelog

Unreleased

  • Fixed: the Add a LoRA and Add a wildcard lists opened off to the side of their button. They now open just under it.

  • Model info: under the model on the generate page, keep where a checkpoint came from and the sampler, scheduler, steps, CFG, size, negative prompt and hires fix its page recommends, then put them on the form with Apply recommended. Kept by file name for the built-in engine, ComfyUI and AUTOMATIC1111 / Forge, in latentry.models.json next to the settings file, for this computer and owner devices only. Latentry never opens the link. Settings → Model info lists every model kept, and adds info for a model that is not loaded.

  • The manual is arranged by what you want to do: Generate is split into the generate page, From a picture (img2img, inpainting, Extend), Bigger pictures (hires fix, Upscale) and Many pictures (the queue, X/Y grid, wildcards). Owner device has its own page under Devices, and checking a release download is on Install from source. The existing pages keep their addresses.

  • Shorter Settings and Generate pages: Settings shows one section at a time, chosen from a list on the left; a dot marks a section with unsaved changes or a field to fix, and /settings#gallery opens a section directly. On Generate, the Show row hides the parts of the form you do not use (characters, source image, pose, hires fix, X/Y grid). A hidden part is not sent, and keeps what it holds until it is shown again. Both choices are remembered in this browser only, and not in secret mode.

  • Fixed: with the wildcard and LoRA pickers beside it, the Positive prompt label broke onto two lines. It now stays on one, and the controls next to it wrap instead.

  • Owner devices: when you do not use a browser on the computer running Latentry (in Docker, or on a computer you use from your laptop), run npm run owner-code in a terminal there and type the code on the pairing page. That browser becomes an owner device: it may change Settings and Setup and pair your other devices, as that computer may. Only a terminal on that computer can make an owner code, so an owner device cannot make another. Settings marks it Owner. Before, the only way was SETTINGS_EDIT=lan, which let every device on the network change settings.

  • An owner device stays one for 7 days, and only a new owner code renews it: pairing again with an ordinary code makes it an ordinary device.
  • Latentry's log records each owner added or removed (the device's number and the time only).
  • npm run forget-devices -- --yes unpairs every device from a terminal on that computer, for an owner device in the wrong hands.
  • Over HTTPS (a reverse proxy in front), the device cookie is marked Secure; the manual shows a Caddy example.

  • Pairing a phone no longer costs a waiting owner code one of its tries.

  • Fixed: after secret mode was turned off, going to the gallery and back could bring back pictures of runs made in secret mode under Earlier. A page that has been in secret mode now keeps none of its runs for that trip, even after secret mode is turned off; a reload starts it afresh.

  • Move a joint in the 3D pose view: drag the head, a shoulder, elbow, wrist, knee, ankle or a foot to change the detected pose. What hangs from the joint moves with it and no bone changes length, so a raised elbow takes the forearm and hand along. Undo takes back the last move, of a joint or of the camera, in the order they were made; Reset joints puts the joints back. Use this pose (before: Use this angle) draws the skeleton with the moved joints and the angle. The moves are kept in the page only. Latentry's engine now needs poseorbit 0.3.1 (run Repair / update on Setup once), which also keeps a stray fingertip's guessed depth from drawing a long line once the pose is turned; for other backends, /api/pose takes edits and answers movable (see the backend API).

  • Accessibility: screen readers now announce what each drop-down list is for, such as Backend, Sampler or Tag spelling. Before, the lists on Generate and Settings had no name. The red "Non-commercial" label on Setup is darker in the light theme, so it is easier to read. The browser tests now check each page with axe-core, in both themes.

  • Fixed: before the engine was installed, a model's Download button on Setup stayed off with nothing saying why. The Models card now says that models are downloaded by the engine, so it has to be installed and started first.

  • Japanese: the UI is now in Japanese too. A browser set to Japanese gets it on its first visit, and the language menu in the header switches between Japanese and English. The manual stays in English. The translation was written without sending the UI's text to a translation service, and a new file translation provider lets others do the same.

  • Support: a new Support page in the manual says which systems, Node.js versions, GPUs and browsers are supported, and which are experimental: macOS on Apple Silicon, and Safari and Firefox. Only the latest release gets fixes.

  • Contributing: a new CONTRIBUTING.md says how to report a problem, set up a checkout, run the checks and open a pull request. The bug and idea forms on GitHub ask you to leave out paths, names and private pictures, and send security reports to private reporting.

  • Compatibility: a new Compatibility page in the manual says what stays compatible between releases: pictures' records, the settings file, the documented environment variables, the backend API (now marked version 1) and page addresses. It also says how a change is announced first. The tests now read a settings file and a picture saved by each release since 0.5.0, so a change that breaks an old one fails before it ships.
  • Docs: the Backends page has a table of what each backend can do: the engine with SDXL or Anima, a diffusers server, A1111 / Forge, and ComfyUI with the built-in graph or your own workflow. Anima's limits (no inpainting or Extend, no hires fix together with a pose) are in it.
  • Changed: a picture saved by a newer Latentry, with a higher record schema, is now read for the fields this release knows. Before, its settings were not read at all. A settings file from a newer Latentry is read the same way, and is no longer saved over, since that would drop what this release does not know. The Settings page says so instead.
  • Fixed: between one queued run ending and the next starting, the button briefly read Generate although a click would still add to the queue. It now reads Add to queue until the queue is empty.

0.12.0 (2026-10-24)

Upgrading from 0.11.0: nothing to do beyond the usual update. The engine is unchanged (still 0.7.0, with the 2026-10-21 lock); the next Install or Repair / update fetches uv 0.13.5 by itself. Updating Node.js to 22.23.4, 24.22.1, 26.12.1 or newer is recommended (see below).

X/Y grid

  • X/Y grid: Generate can draw a run again for each value of one or two settings (CFG, steps, seed, hires denoise, sampler, scheduler, or a prompt search and replace, which also compares LoRA weights), all on one seed, and lay the pictures out as a table. The grid waits in the queue as one run, its cells run with nothing in between, and Cancel grid stops it whole. Each picture records its place in the grid. The gallery can stack a grid's pictures, and its X/Y grid filter shows only grids' cells, only their sheets, or only other pictures. Save sheet saves the table as one picture when asked (never in secret mode).
  • Fixed: pictures saved in the same second with the same seed could overwrite each other. Names are now reserved until the picture is written.

Security

  • The engine install uses uv 0.13.5 instead of 0.13.0. uv before 0.13.4 trusted the emSign ECC Root CA - G3 certificate authority, some of whose intermediate certificates had their keys compromised, so a forged certificate could have been used to intercept the packages it downloads (GHSA-hq85-5cpj-mr62). 0.13.5 drops that root and also rejects chains through the compromised intermediates. uv runs only during Install and Repair / update, and each of those fetches the pinned uv first and checks its SHA-256, so the next one moves an existing engine to 0.13.5 with no extra step. The engine's own Python already uses certifi 2026.10.19, which drops the same root.
  • Latentry stops trusting the compromised emSign roots on any Node.js. Its own downloads (models, taggers, uv) go through Node.js, and Node.js dropped emSign ECC Root CA - G3, emSign Root CA - C1 and emSign ECC Root CA - C3 only in 22.23.4, 24.22.1 and 26.12.1 (2026-10-21). On an older Node.js the server now takes those three out of what it trusts when it starts, and says so in its log. Updating Node.js is still recommended.

Dependencies

  • Next.js 16.4.3 instead of 16.4.2. Three published Next.js advisories (a Server Actions denial of service, and remote code execution through the image optimizer with SVG files and through use cache) affect 16.0.0 to 16.4.0; 16.4.2 was already outside them, and 16.4.3 is the newest patch.
  • The tests also run on Node.js 26.

0.11.0 (2026-10-22)

Upgrading from 0.10.0: after the usual update, open Setup and run Repair / update for the engine once. It installs the engine's Python packages from the new lock (resolved for 2026-10-21), with Pillow 13.0.0, which fixes known vulnerabilities in Pillow 12.3.0. Setup then says the packages are pinned as of 2026-10-21. The engine itself is unchanged (still 0.7.0).

The queue

  • A queue for each backend: while a run is going, Generate becomes Add to queue, and the run starts as soon as the one before it ends (up to 8 waiting on each backend). Waiting under the results lists the runs this page added, and any of them can be taken out again. Other pages and devices see only how many runs are waiting. A web UI or ComfyUI that is busy with another client's run is waited for instead of refused. The queue is held in memory only, and a second click no longer just shows the run already going.
  • Earlier runs stay on Generate: when a new run starts, the pictures on screen move down to Earlier instead of vanishing, newest first, for the last 5 runs (32 pictures). They and the queued runs are held in the browser tab's memory only: going to the gallery and back keeps them, a reload clears them, and in secret mode they last only while Generate is open. Each run folds from its heading, which shows the whole prompt on hover; in secret mode no prompt is shown there.

Dependencies

  • The engine's Python packages are pinned as of 2026-10-21: Pillow 13.0.0 (fixing eight vulnerabilities reported against 12.3.0, in reading PDF, PPM and other files), with newer fastapi, starlette, transformers and pydantic.
  • Next.js 16.4.2, React 19.3.1, Radix UI 1.8 and other web packages are updated, as are the tools that check and document Latentry.

0.10.0 (2026-10-20)

Upgrading from 0.9.0: nothing to do beyond the usual update. The engine is unchanged (still 0.7.0).

Generate

  • Extend a picture (outpainting): a new Extend mode on the source image grows the canvas by up to 1024 pixels on each side and draws the new area, with a preview of the new canvas. It works on every backend that can inpaint: Latentry pads the picture and makes the mask itself, so the backend gets an ordinary inpaint.
  • Wildcards: __name__ in a prompt takes one line of name.txt from a wildcards folder set on Settings, once per run (the same line again for the same fixed seed). __folder/name__ reads from a subfolder, as in A1111's Dynamic Prompts extension, and Add a wildcard next to the prompt puts one in. The picture keeps the prompt as typed too. Paired devices can use wildcards only once Use on paired devices is on.
  • Screen readers now hear a name for every slider on Generate, for the gallery's filter menus and for the saved-characters menu.

Manual and tests

  • The guide checked against the app: browser tests now walk through what the guide says about Generate, the gallery, pose, secret mode, Setup and pairing. Where they disagreed, the guide was corrected: a secret run's pictures are saved to the gallery (the threat model said they were not), and Settings now describes the folders' display names and the pairing card, and the guide's Setup page the switch for the LoRA list on paired devices.
  • New screenshots in the guide: the gallery and its viewer as they look now, filters, Compare, hires fix, upscaling a source, secret mode on Generate, pairing a phone, the 3D view, Setup with the engine running and a ComfyUI workflow; the README's gallery and viewer shots are retaken too. npm run docs:screenshots retakes them from a clean Latentry, with public example pictures or a gallery the engine draws for the purpose.
  • Runs of several pictures are tested: browser and route tests now draw two pictures in one run on an A1111 web UI and on a diffusers-compatible server, and check that each is shown, seeded one after the other, and saved. The limit of 16 pictures a run is one shared setting in the code.

Fixed

  • The untouched part of a ComfyUI inpaint: inpainting and Extend on ComfyUI left the part of the picture that was meant to stay slightly changed, because it went through the VAE once. The built-in graph now pastes the drawn area over the source's own pixels, as Latentry's engine and A1111 do.
  • A prompt typed just before switching backends was lost from the first backend's form.
  • The saved-characters box went blank once secret mode was turned off, instead of saying there are none.

0.9.0 (2026-10-18)

Upgrading from 0.8.2: after the usual update, stop and start the engine on Setup once, so it runs engine 0.7.0 with hires fix and Upscale (Setup says so while the old one runs). There is nothing new to install.

Hires fix

  • A Hires fix section on Generate: each picture drawn from the prompt (with or without a skeleton pose) is upscaled and redrawn at the larger size: choose the upscaler, how much larger, the hires steps (0: as many as the first pass) and the denoising strength. Off by default. The progress bar counts both passes, and the second pass runs exactly the hires steps on every backend.
  • On A1111 / Forge it is the web UI's own hires fix, with its upscalers (Lanczos, R-ESRGAN and the like) and latent modes. Only upscalers already on disk are offered, so none makes the web UI download a model. Forge and its forks get the fields their second pass needs and the first pass's CFG.
  • On ComfyUI the built-in graph upscales with Lanczos or an upscale model from ComfyUI's upscale_models folder (listed by file name, so its folders stay on that machine), then samples again with the same prompts, LoRAs and skeleton. A workflow from the folder keeps its own passes.
  • On Latentry's engine SDXL and Anima models upscale with Lanczos (no upscale model to fetch) and redraw on the same weights and LoRAs, one picture at a time, decoding the larger picture in tiles. With SDXL the second pass follows the skeleton too; on Anima, hires fix and a pose do not go together yet.
  • Saved like A1111: the picture records the size it was drawn at with Hires upscale, Hires steps and Hires upscaler, so other tools read it, and the gallery's Generate with these settings turns hires fix back on. Reading an A1111 picture no longer takes its hires Denoising strength for an img2img source's.

Upscale

  • Upscale a picture from the gallery: a new Upscale button in the side panel (and a third mode of the source image) makes a finished picture larger and redraws its detail, hires fix for a picture that already exists. The picture's own prompts come along. A1111 / Forge enlarge it in the Extras tab with the chosen upscaler, ComfyUI in the graph with Lanczos or an upscale model, Latentry's engine with Lanczos; then one redraw at the hires denoise. Up to 4096 pixels a side.
  • Recorded as an upscale: the saved picture says mode upscale, the size it had before and its hires settings.

Gallery

  • The real size: the side panel's Size is now the picture's own size. For a picture drawn smaller and then enlarged, a Drawn at line shows the size its settings name. The panel also shows a picture's hires settings.

Backend API

  • /api/health may answer hires, upscale and upscalers, and /api/generate takes hr_scale, hr_steps, hr_denoise, hr_upscaler and upscale, for diffusers-compatible servers that want to offer them (see docs/backend-api.md).

0.8.2 (2026-10-14)

Upgrading from 0.8.1: nothing to do beyond the usual update. On Windows, the next start gives Latentry's private files their owner-only access list.

Security and privacy

  • Private files restricted at start: on Windows the settings file, the pairing key, the paired-device list and the folder-name key now all get their owner-only access list when Latentry starts. Before, each got it the first time it was used, so a pairing key nobody had used since an update stayed readable by other users until then.
  • The version on Settings: Settings shows the version of Latentry it runs, so it can be matched against published security advisories. Like the rest of Settings, it is shown on this computer only. The LIGSP conformance page gets the profile's new OPS-3 row.
  • The engine's version on Setup: each running engine shows its version, and Setup says when it differs from the engine in Latentry's folder, which means it was started before an update and needs a stop and start.

Manual

  • LIGSP conformance: the manual has a new page with Latentry's conformance table for the Local Image Generation Security Profile (LIGSP 0.1, level 2), the profile Latentry is the reference implementation of. The README, SECURITY.md and the threat-model page link to it, and the threat-model page says how to check your own setup with the profile's read-only checker (uvx local-image-gen-security check ...).
  • Checking a download against its tag: Getting started now adds --source-ref refs/tags/vX.Y.Z to gh attestation verify, which also checks that the file was built from that release's tag.

0.8.1 (2026-10-13)

Upgrading from 0.8.0: nothing to do. The first load of a single-file SDXL checkpoint fetches SDXL's config files once (a few MB of text) into the engine's controls folder. A diffusers folder model whose weights are only in a pickle format is now refused; get its .safetensors files.

Security and privacy

  • Settings says whether Latentry is open to your network: the pairing card now says when Latentry listens beyond this computer (LATENTRY_HOST), and how to open it when it does not.
  • Private files stay private on Windows: the settings file, the pairing key, the paired-device list and the folder-name key now get an access list with only your account (and Windows itself), so other users of the computer cannot read them wherever Latentry is installed. Files from earlier versions are fixed the first time they are read.
  • What Latentry protects against: a new page in the manual goes through who could try to misuse Latentry (a website you visit, another device on your network, internet scanners, model files and workflows, swapped downloads, someone using the computer later, whoever receives a picture), what Latentry does about each, and what it leaves to you. SECURITY.md and the network guide now say that paired devices share ordinary runs.
  • Checked release downloads: each release now has its source as .tar.gz and .zip, made by a GitHub workflow from the release's commit, with SHA256SUMS and a build provenance attestation, so gh attestation verify can show where a download came from. Getting started says how to check one.

Engine (0.6.1)

  • The engine reads only safetensors: a diffusers folder model is now refused when a part of it has weights only in a pickle format (.bin, .ckpt, .pt), which can run code when loaded, and the message names that part. The SDXL folder and ControlNet loads also ask for safetensors only. Single-file checkpoints and LoRAs were .safetensors only already.
  • Pinned SDXL configs: a single-file SDXL checkpoint's config files (scheduler, text encoders and tokenizers; no weights) now come from a fixed commit of SDXL base, each checked against its SHA-256, and are kept in the engine's controls folder. Before, diffusers fetched them from the latest version on the first load. An inpainting checkpoint is loaded as before.

0.8.0 (2026-10-13)

Upgrading from 0.7.0: open Setup and run Repair / update for the engine once. It installs the engine's Python packages from the new lock and brings poseorbit 0.2.0 (picking any of the people in a pose picture).

Pose

  • Pose on ComfyUI: a ComfyUI backend now follows a skeleton pose, with the Pose ControlNet chosen for it on Settings (or GEN_CONTROLNET_<ID>), an OpenPose ControlNet that fits the checkpoint (NoobAI-XL OpenPose for Illustrious and NoobAI models, xinsir's for other SDXL models; the manual's Backends page says why). Latentry's engine finds the people and draws the skeleton, so picking people and the 3D view work as on the engine, and only the drawn skeleton goes to ComfyUI, never the reference picture. It works with a source picture and a mask too; a secret run blanks the uploaded skeleton with the rest.
  • Pose in your own ComfyUI workflows: a workflow with a Load Image node titled latentry:pose follows the skeleton pose through its own pose nodes (an Anima pose adapter's, say). The skeleton is drawn dwpose-style unless the title also says openpose, and a node titled latentry:pose_strength takes Pose strength. The Workflow choice says which workflows can follow a skeleton.
  • Pick any of the people for a pose: in a picture of several people, click each numbered box to follow that person or leave them out, so two of three can be used; at least one stays. Everyone still picks all. This needs a backend that says it takes a list (person_subsets in docs/backend-api.md); Latentry's engine does from poseorbit 0.2.0, so run Repair / update for the engine once. Other backends keep one person or everyone, and the chooser says why.

Generate page

  • Sliders stay put while you scroll on a phone: on a touch screen a slider now moves only when you drag its knob, which has a larger grip. A swipe that starts on a slider's track scrolls the page, and a tap on the track no longer changes the value. A mouse click on the track still moves it there.

Backends

  • A warning for backends open to the network: Settings now says, under the backend, when a ComfyUI or AUTOMATIC1111 / Forge backend is reached at another computer's address without a token. Anyone who can reach such a server can use it without Latentry (ComfyUI has no login, and its custom nodes run code). The manual's Backends on other computers says how to keep one on this computer, or put a login in front of it.
  • Forge Neo: checked as an a1111 backend (Forge Neo 2.30, started with --api): txt2img, img2img, inpainting, progress, Cancel, samplers and schedulers, and its pictures in the gallery.

Engine install

  • Pinned Python packages: the engine's packages, PyTorch included, now come from a lock for your machine and PyTorch build (engine/locks/), every file checked against its SHA-256, and Repair / update makes the environment match it exactly (a package added to it by hand is removed). Setup shows the date they were pinned and warns once that is over six months ago. Where there is no lock, or with LATENTRY_ENGINE_PACKAGES=latest, the engine is installed unpinned as before, and Setup says so. Run Repair / update once to switch.
  • PyTorch for 570-series NVIDIA drivers: GPUs older than Blackwell now get the CUDA 12.6 build, which PyTorch still updates, instead of CUDA 12.8, which stopped at an older PyTorch. Where an older build is the only one that runs (CUDA 11.8, or Blackwell on a 570 driver), Setup suggests a driver update.
  • Pinned pose weights: the engine now fetches the SDXL OpenPose ControlNet and the Anima pose adapter from a fixed commit and checks each file against its SHA-256 before using it. A file fetched before this that does not match is fetched again once.
  • Pinned pose detectors: the engine needs poseorbit 0.1.2, which does the same for its three detector models and no longer leaves a second copy in Hugging Face's shared cache. Run Repair / update for the engine on Setup once to get it.

Engine (0.6.0)

  • Faster runs with the same LoRA: the engine leaves the last four LoRAs used in the model, switched off, so a run that calls one again skips putting it in (about 1.5 s on SDXL). A run without them comes out exactly as before. LATENTRY_LORA_KEEP sets how many; 0 takes each out after its run, as before.

Fixed

  • Progress on A1111 and Forge after a change of steps: an img2img or inpaint run right after a run with another step count showed the earlier run's count, so its bar stopped short or ran past the end. The count is now taken again once the web UI sets it.

0.7.0 (2026-10-11)

Upgrading from 0.6.0: open Setup and run Repair / update for the engine once. It moves the engine onto its own Python and the pinned uv (about 20 MB more, and PyTorch may download again). Network folders (\\server\share) in the settings file are dropped unless LATENTRY_ALLOW_UNC=1 is set.

ComfyUI

  • ComfyUI as a backend: add a ComfyUI server on Settings with the kind comfyui. Latentry builds the graph itself (checkpoint, LoRAs from <lora:…> in the prompt, KSampler), shows progress, and Cancel stops only its own prompt. Choose the checkpoint under Model after Test, or with GEN_MODEL_<ID> in .env.local; without one, ComfyUI's first is used. Pose is not available on ComfyUI yet.
  • img2img and inpainting on ComfyUI: a ComfyUI backend takes a source picture and a mask. They are uploaded into ComfyUI's input folder under two names per backend that each run reuses.
  • Your own ComfyUI workflows: give a ComfyUI backend a Workflows folder on Settings, and the Generate page offers the workflows saved there (ComfyUI's Export (API)) next to the built-in graph. Latentry fills in the prompts, seed, size, steps, CFG, sampler and LoRAs where it finds them, or where a node's title says latentry:prompt and the like, and says what it could not find. The backend's Model, if set, replaces a workflow's checkpoint; a workflow with a UNET loader keeps its own model. Workflow names reach paired devices only when you turn that on.
  • What stays on the ComfyUI computer: pictures go to ComfyUI's temp folder rather than its output folder, and the copy Latentry saves leaves out ComfyUI's graph, which names your model folders. After a secret run, Latentry deletes its prompts from ComfyUI's history and overwrites its uploads with an empty picture.
  • More from ComfyUI pictures: the gallery reads the settings of more ComfyUI graphs: hires and refiner graphs (the base pass), SamplerCustom and Flux-style graphs, other model loaders, seeds from a seed node, and pictures that carry only the editor's workflow. Model names show the file name only, not the folders it sits in.

Secret mode, privacy and security

  • Secret mode turned on during a run: the run becomes secret from then on (within a minute of its end too): a reload or another device no longer picks it up, and it is forgotten a minute after it ends. Turning secret mode off again does not undo that.
  • Less of this computer on paired devices:
  • Gallery folders are shown by a display name you set next to each folder on Settings, or else by a made-up name such as "Amber Fox". The folder's own name is no longer shown, here or on a phone.
  • Unexpected errors from the engine no longer pass on their own text, which can name files; it goes to the engine's log.
  • Paths in other errors (the built-in tagger, other backends) show as <path>.
  • Model names show the file name only, without the folders they sit in.
  • npm run dev warns when it listens beyond this computer, since its error pages show paths and source code.
  • LoRA list on paired devices: Show the LoRA list on paired devices (Setup > Engine, off by default) gives phones and other paired devices Add a LoRA too.
  • Tighter API limits: network folders (\\server\share) are refused on the Settings page unless LATENTRY_ALLOW_UNC=1, and are dropped from an existing settings file without it. A backend's redirect is reported as an error instead of followed. Requests without a picture are capped at 1 MB. Tagging and pose previews run at most two at a time. Pages are no longer kept in the browser cache and get stricter security headers. The unused /_next/image endpoint is off.

Backends

  • Why a backend does not answer: Test on Settings and the Generate page now say whether a backend asked for a login (for example a ComfyUI login custom node, or a web UI started with --api-auth and no token), answered an error, redirected, or did not answer at all, with its own message, instead of only "No answer". For every kind of backend.

Engine install

  • Pinned downloads: the engine install fetches uv 0.13.0 instead of the latest release, and checks it against its SHA-256 before running it. An engine installed with another uv gets 0.13.0 on its next Repair / update, which may download PyTorch again once (uv 0.13 changed its cache format). Models and taggers from the Setup page come from a fixed commit of their repository, and each file is checked against its SHA-256. A file that does not match is not kept, and the page says so.
  • The engine keeps its own Python: the engine install downloads Python into .runtime/ instead of building on a Python 3.12 already on this computer, so updating or removing that Python no longer breaks the engine. Your own uv settings (UV_* variables, uv.toml) no longer change how the engine is installed. An existing engine is moved onto its own Python at its next Repair / update (about 20 MB more, and its packages are installed again).

Engine (0.5.0)

  • Stricter API: the built-in engine caps request bodies at 64 MB and prompts at 20,000 characters, checks /api/pose requests field by field, takes only owner/name repository ids for downloads, keeps a short download list, and no longer serves FastAPI's /docs pages.
  • Model folders downloaded from the Setup catalog are checked against their pinned hashes, and unexpected errors go to its log instead of to the client.

Fixed

  • A web UI backend moved to another address on Settings shows that server's samplers at once, not the old one's for a minute.
  • Pictures from ComfyUI that carry only the editor's workflow are read with an SDXL text encoder too.

0.6.0 (2026-10-10)

LoRAs

  • LoRAs on Latentry's engine: put LoRA files in the loras folder inside the models folder and call them as <lora:name:0.8>, as on A1111, or pick one from Add a LoRA next to the prompt. SDXL and Anima LoRAs are both supported. They are on for that run only. A missing LoRA, or one made for the other model family, stops the run with a message.
  • Write LoRA hashes (Settings > Gallery folders, off by default): saved pictures made with LoRAs carry A1111's "Lora hashes". These identify the exact files, so leave it off unless you want sites to link the LoRAs.
  • LoRA filter in the gallery: Filters has a LoRA menu listing the folder's LoRAs, and the search takes lora:name and -lora:name. Both go by the LoRAs a picture was made with.
  • LoRAs in the viewer: the gallery lists a picture's LoRAs only when the backend applied them. A <lora:…> call that Latentry's engine or another diffusers server read as plain words is no longer listed, in new pictures or old ones. Pictures from A1111 are listed as before.

Viewer

  • The secret-mode eye and the blur in the picture viewer now come from the viewer library itself (panorail 0.3.0) instead of being added to its page afterwards, so they keep working when the library changes. They look and work as before.

Engine (0.4.0)

  • Loads LoRAs from the models folder's loras folder for a single run, alongside the pose adapter, and lists them at GET /api/loras. Its health reports the lora capability, which the generate page reads.

Fixed

  • With npm run dev, the first page opened on this machine after a start went to the pairing page. Reloading worked. Requests are now known to come from this machine from the very first one.
  • A prompt that names the same tag twice no longer causes an error in the viewer's tag panel, and choosing one of the two opens only its own actions.

0.5.0 (2026-10-08)

Upgrading from 0.4.0: phones and other computers must now be paired before they can use Latentry (this computer needs nothing). After updating, open Settings > Phones and other computers > Add a device on this computer and type the code on each device. Behind a reverse proxy that does its own login, set LATENTRY_PAIRING=off.

Pairing and security

  • Pairing: a phone or another computer must now be paired before it can use Latentry. On this computer, Settings > Phones and other computers > Add a device shows a code (one use, 10 minutes) to type on the other device, with a name if you like; it then stays paired for 30 days. The same card lists the paired devices with when each ends, and removes one or all of them. A paired device's Settings page shows when its pairing ends and can unpair it, and every page warns from three days before. The list keeps no record of use and is stored encrypted. The key can be kept in the system keychain and passed in LATENTRY_PAIRING_KEY; the network guide shows how on Windows, macOS and Linux. LATENTRY_PAIRING=off turns it off, for a reverse proxy that does its own login.
  • The per-device request limits count the address a connection really comes from, and believe X-Forwarded-For only from this computer. Stopping a run takes that run's id.

Secret mode

  • Every picture is blurred until you choose to show it: the gallery's viewer (one picture at a time, with Show for this picture, blurred again on the next one or when it closes), the generate page's results and their full-size view (Show results, per run), the source picture in its preview and the mask editor (strokes stay sharp), and the picture you pick a pose's person from (its numbered boxes stay sharp). On the generate page each has a Show / Blur button, so a picture can be blurred again, on a phone too. The full-size viewer has its own secret-mode switch in the bar, so it can be turned on without closing the picture; turning it on blurs anything shown before.

Generate

  • A slider moved away from its default shows a reset arrow next to its label, which puts it back to the profile's value. This covers width, height, steps, image count, CFG and both strengths.
  • On a backend that does not load LoRAs (anything but an A1111 web UI), a <lora:…> or <lyco:…> tag in the prompt shows a note that it is read as plain words.
  • From gallery (source and pose pictures) opens in place, under its button, instead of as a dialog over the page, so the header and its secret-mode switch stay one tap away while you choose.

Engine (0.3.0)

  • The engine answers only requests addressed to this computer, and run by hand it refuses to listen beyond it without LATENTRY_ENGINE_TOKEN. The token is compared in constant time.
  • The engine has tests (models folder, GPU plan, the HTTP API's token and input checks), which CI runs without a GPU or a model.

Fixed

  • img2img with a detailed source picture and a mask no longer fails with "Invalid JSON body". Latentry accepted only the first 10MB of a request and cut the rest. It now takes up to 48MB, and above that the page says the pictures are too large before sending them.

0.4.0 (2026-10-08)

Upgrading from 0.3.0: Latentry now needs Node.js 22.19 or newer (was 22.12), for undici 8. Update Node.js first, then update Latentry as usual.

Gallery

  • Order pictures oldest first, by name, by file size, by when they were made or by pixel count. Filter them by shape, file type, age, resolution, model, where the settings came from, and whether they have WD14 tags.
  • The search box takes settings as key:value (model:noobai, steps:>=30, w:1024), and leaves pictures out with a leading - (-smile, -"long hair", -model:pony).
  • Stack pictures that share a prompt or a seed into one card, and open a stack to see them. Compare two selected pictures with a slider (or side by side), their differing settings and the prompt tags only one has.
  • Settings are read from JPEG and WebP too (A1111 / Forge's EXIF, ComfyUI's WebP), along with the LoRAs a picture used, which the viewer lists. Before, only PNGs had settings. Only the headers are read, and from EXIF only the fields that hold settings.
  • Show in folder and Open in default app in the viewer open a picture in this computer's file manager (selected) or its picture app. They appear only when the browser runs on the same computer, and never act on a file outside the gallery folders. The gallery still has no delete, move or rename, by design.
  • The order and filters are not kept in the browser; what the gallery reads to sort a folder stays in memory.

Privacy

  • A finished run's pictures and prompt leave Latentry's memory once no page may load them any more (an hour, a minute for a secret run), instead of staying until that backend's next run.
  • The engine reads a single-file SDXL model's config files from Hugging Face's cache after the first load, instead of asking the Hub each time.
  • Prompt fields and search boxes no longer use the browser's spell check (which some browsers do online) or autofill.
  • Secret mode: a character saved while it is on is kept in that tab, marked temporary, and dropped when secret mode is turned off (before, it vanished on the next reload). A comparison shown with Show is hidden again when it is closed.
  • A new Privacy page in the manual lists what goes over the network and what is kept where, and the Secret mode page covers Compare, Download and what turning it off keeps.

Backends and install

  • When a backend cannot use the source image, inpaint mask or skeleton pose, the form says why and what would turn it on, instead of just leaving it out. The same goes for a pose without the 3D view, and for a skeleton dropped after switching to a backend without pose. Servers can give their own reason in a new optional /api/health field, unavailable.
  • install.sh --no-engine (install.ps1 -NoEngine) is for servers you already run: it turns off the built-in engine in a new .env.local and points you to Settings instead of the Setup page.

Look

  • README and manual lead with what Latentry is for: private, local image generation, and a gallery that remembers how every picture was made. A new banner, also drawn by the logo that the install scripts and npm start print, which fits a 120-column terminal.
  • Dark mode: the 3D pose view and the skeleton preview have a visible edge, dialogs dim the page behind them, and sliders show their whole track. The 3D view's floor grid is a little brighter in both themes.

Fixed

  • Viewing a WebP or JPEG in the gallery no longer keeps the file open, so it can be renamed or deleted elsewhere while Latentry runs.

0.3.0 (2026-10-07)

Upgrading from 0.2.0: a backend token is now only sent to the server it was set up for. A GEN_TOKEN_<ID> applies only while GEN_BACKENDS lists that id at the same address (scheme, host and port); a backend added only on the Settings page needs its token typed there. Moving a backend to another address on Settings asks for its token again.

Security

  • Backend tokens could be sent to another address by anyone who can change settings: Test lent a saved token to any URL, and a backend kept its token when its address changed. Tokens now stay with their server.
  • "Only from this machine" (the default SETTINGS_EDIT=local) went by the Host and X-Forwarded-For headers, which a program on the LAN can write. It now also checks the address the connection really came from.
  • npm start, npm run dev and npm run build turn Next.js's anonymous usage reports off; the engine runs with Hugging Face's off and without the other backends' tokens in its environment.

Secret mode

  • Easy to see: while it is on, the header shows a filled Secret pill.
  • A run started in secret mode stays on the page that started it: no other tab, reload or device picks it up, and the server forgets it a minute after it ends.
  • What the gallery sends to the form is no longer written to browser storage; the gallery picker blurs pictures and WD14 tags are hidden too.
  • In the gallery panel, Show for this picture reveals one picture's prompt, settings and tags while it is on screen, and analysing a picture says how many tags were found while they are hidden.
  • A new manual page says what it hides, and what it does not: pictures are still saved to the gallery folder with their prompt, and shown in the gallery.

Generate

  • Prompts are sent in the tag spelling of the profile's model family: underscores for SDXL, Illustrious / NoobAI and Pony, spaces for Anima, as typed for generic. Settings or tags sent from the gallery arrive in that spelling. Each profile's spelling, and the tags it never respells (score_* by default), can be changed on Settings.
  • Share across backends next to the prompt keeps one prompt and artist for every backend; the negative prompt stays per backend.
  • long_hair and long hair count as the same tag when tags are added to a prompt.

Gallery

  • Pictures and thumbnails are no longer kept in the browser's cache.
  • A card's full tag list folds back with Show less (bigger on touch screens), a click or tap outside the card, or Esc.

Fixed

  • SDXL img2img counted one step too many whenever steps × strength was not a whole number, so a batch stopped at e.g. 23/24 until its pictures came.
  • Sending a Pony picture's settings to the form no longer turns score_9 into score 9.

0.2.0 (2026-10-06)

First public release. The manual: https://33rd-kk.github.io/latentry/

  • Generate: text-to-image, img2img (variation, or a picture as a pose reference), inpainting with a painted mask, one form per backend with model profiles (Anima, SDXL, Illustrious / NoobAI, Pony, generic).
  • Pose: pick whose pose in a group picture, turn it in 3D and frame it (zoom, move) in a view that shows exactly the skeleton the run will follow (poseorbit).
  • Built-in engine: installed and run from the Setup page, one per NVIDIA GPU, stock diffusers; SDXL with an OpenPose ControlNet and Anima with its pose adapter.
  • Backends you already run: any server speaking the backend API, and AUTOMATIC1111 / Forge.
  • Gallery: search by prompt, tags, model and file name; WD14 tagging on the CPU; send a picture's settings or tags back to the form.
  • Safety: listens on this computer only unless LATENTRY_HOST opens it to the network; same-origin and private-network-only /api, per-IP budgets, settings editable from this machine only by default; gallery pictures are read through the file that was checked to be in its folder.
  • Install: install.sh / install.ps1 show the Latentry logo, build, and say which address to open; pose detection comes from poseorbit on PyPI.