What changed in Locally Uncensored 3.0.1?

Qwen-Image 2.1 is downloadable in the Model Manager and runs locally on Windows and Linux, generating from a prompt and editing from a reference image. Models has a LoRAs tab of its own with CivitAI search, an installed list and deletion. A GPU whose free memory could not be measured is planned with a bigger safety margin instead of as an empty card. The Linux AppImage no longer hands its own runtime to the programs LU starts. The local model answers one conversation at a time and says how many are waiting.

What does Qwen-Image 2.1 need?

About 16 GB across three files, and ComfyUI 0.37.0 or newer for the encode node it runs on. An older ComfyUI is told which version to update to before anything is built. It takes one reference image at a time and no painted masks, because that encode node accepts neither. The weights carry the Qwen Research License, which allows research and evaluation but not commercial use, and the bundle says so with a link before the download starts.

Why did a Linux install break a working Python before 3.0.1?

The AppImage runtime exports its own LD_LIBRARY_PATH, PYTHONHOME and related variables, and every program LU started inherited them: git, a system Python, pip, ffmpeg, nvidia-smi, the coding agent shell and the Character Trainer. A perfectly healthy system Python then failed to import ssl or find its standard library, and the error pointed at the Python install rather than at the inheritance. From 3.0.1 those variables stop at the AppImage boundary.

Is there a macOS build of 3.0.1?

No. Every release builds Windows and Linux only. On a Mac, LU Cloud runs in the browser at lu-labs.ai.

Why does a chat sometimes wait its turn now?

The local model runs one conversation at a time. Before 3.0.1 a second chat sent while the first was still streaming could mix the two answers together. A chat that has to wait now says so, with a line showing how many are ahead of it, and Stop works while it waits and takes it out of the line.

v3.0.1: One Local Model That Draws and Edits, and a Calmer Linux

Locally Uncensored 3.0.1 follows 3.0.0 with two additions and a long list of corrections. The additions are a local image model that generates and edits with the same weights, and a place in Models where your LoRAs finally live. The corrections are mostly about machines that were fine and were treated as if they were not: a graphics card that cannot report its free memory, a system Python that a Linux install made look broken, a processor without AVX2.

Quick TL;DR before we dig in:

Qwen-Image 2.1 downloads in the Model Manager and runs on your own card: prompt to picture, or reference image plus prompt to edit, with no mask to paint
LoRAs have their own tab in Models, with CivitAI search, an installed list with sizes and trigger words, and deletion
A card whose free memory could not be measured is planned with a bigger margin, and the log stops calling total capacity free
The Linux AppImage keeps its runtime to itself, so git, Python, pip and ffmpeg behave the way they do outside the app
One local conversation at a time, with a queue that says how many are ahead and a Stop that works while waiting
Sampling belongs to the chat you moved it in, not to every open chat at once

Qwen-Image 2.1, Generating and Editing

Qwen-Image 2.1 is in the Model Manager and runs locally on Windows and Linux. One model does both jobs. Write a prompt and it generates a picture. Add a reference image plus a prompt and it edits that picture, with no mask to paint first. It shows up in the Generate picker and in the Edit picker like every other installed image model, so nothing about the Create tab has to be learned again.

What it costs you in disk and setup: the bundle pulls three files, about 16 GB together, and it needs ComfyUI 0.37.0 or newer for the encode node it runs on. An older ComfyUI is told which version to update to before anything is built, instead of failing halfway through a graph.

What it does not do, stated up front

One reference image at a time, and no painted masks. Its encode node takes neither, so both are out of reach rather than merely unfinished. The weights carry the Qwen Research License, which allows research and evaluation but not commercial use, and the bundle says so with a link before you start the download.

LoRAs Have a Tab of Their Own

Until now a LoRA sat unnamed among the checkpoints in the Image tab, which is a poor place to look for one. Models now has a LoRAs tab on Windows and Linux. Get searches CivitAI and downloads what you pick straight into ComfyUI's models/loras folder. Installed lists what is actually on disk with its size, marks a character you trained yourself with its trigger word, and deletes a file you no longer want.

There is a matching fix in Create: a character LoRA now offers only the models that were trained for that character's family. The picker could previously suggest a pairing that was guaranteed to fail at generation time, which is a cost you only discovered after waiting for it.

This tab is not offered on a Mac. Local media there runs on Apple MLX, and there is no ComfyUI models/loras folder to list.

A Card That Cannot Report Free Memory

This is the fix the release is named for in the changelog. When free VRAM could not actually be measured, with no nvidia-smi present for instance, the planner used to proceed as though the entire card were sitting empty, and the log line said "N MiB are free" about a number that was really the total size of the card, every other program's usage included.

Two things changed. That weaker reading now gets a bigger safety margin, so a start places fewer layers rather than too many. And the log line says total capacity, because that is what the number is. A start that plans too many layers does not fail politely; it takes the engine down with it.

Two neighbours in the same area: the local engine picks its processor code path at startup, so an older processor without AVX2 runs local models instead of the engine exiting right after start, and when it genuinely cannot run it names the missing instruction set, measured from the processor rather than guessed, and stops retrying a binary that would fail the same way twice.

Linux: the AppImage Stops Leaking Its Runtime

An AppImage runtime exports its own LD_LIBRARY_PATH, PYTHONHOME and related variables, and everything LU started inherited them: git, a system Python, pip, ffmpeg, nvidia-smi, the coding agent shell, and every program the Character Trainer launches. The result was a healthy system Python that could not import ssl or find its own standard library, with a diagnosis that pointed at a broken Python install rather than at the inheritance. Those variables now stop at the boundary.

The distributions that refuse to let pip write into the system Python (Arch, Debian 12 and newer, Fedora 38 and newer, Ubuntu 23.04 and newer) had a second version of the same problem: the isolated environment LU had already built there was killed at the first pip call. That is fixed, and the same fix keeps the Coding Agent's own terminal from inheriting the broken environment for every git, pip or python command typed into it.

Two more Linux corrections worth naming. Installing or repairing ComfyUI now searches the interpreters already on your machine for one that PyTorch actually ships wheels for, and uses it, with no picker in Settings; if there is none it says so and names what to install before starting the roughly 2 GB PyTorch download instead of failing at the end of it. And a ComfyUI install missing python3-venv now says so, rather than ending with "venv creation failed:" and nothing after the colon, which is what happened while LU read only stderr and Python printed its hint on stdout.

Chats Wait Their Turn, and Say So

Sending in one conversation while another was still streaming used to mix their text together, and a second agent run could be dropped silently while the first was going. The local model runs one conversation at a time now, and a chat that has to wait says so, with a line showing how many chats are ahead of it. Stop works while it is still waiting and takes it out of the line. Stop, signing out and quitting reach every conversation, including one that has not started yet, not only the one open on screen.

Alongside that, the sampling controls were quietly shared before: moving Temperature, Top P or Max tokens changed every open chat, because they all read one value from the Settings page. Each now belongs to the conversation you moved it in. A chat with no slider of its own still follows Settings, and a field nobody moved anywhere is left out of the request entirely, so the model applies its own default rather than ours.

Agent and Workflow Corrections

Workflow steps go through the same tool approval as the rest of Agent mode, and only offer the tools your permissions allow. Built-in workflows take their input from the call itself instead of waiting forever for an answer nobody could give, and the Play button under Agent Workflows that never did anything is gone
A sub-agent delegated in the foreground is no longer cut off after 60 seconds, and a timeout or a Stop ends the tool call it was running instead of leaving it going in the background
Pressing Stop while the Coding Agent is running code stops that run, the way it already stopped a shell command
Troubleshoot tests the LM Studio address you configured, instead of always trying the default 127.0.0.1:1234
Picking a specific GPU keeps that physical card for a local model or the Character Trainer, even if the system renumbers the cards between detection and start
Replacing an OpenAI compatible backend parks the displaced API key in the OS keychain instead of dropping it, and gives it back if you switch back. The warning that a key will be lost now appears only on a device with no keychain to park it in

Cloud, Briefly

Two small things on the cloud track. A video render now offers exactly the clip lengths the picked model supports, read from the live catalog, instead of the same fixed pair for every model. And a render started without a prompt, a presenter reading a script for example, gets a real name in the gallery instead of an empty tile. The catalogue itself is unchanged in this release; the counts on the cloud page still hold.

Download

3.0.1 is on GitHub Releases, with one installer for Windows and three packages for Linux. The signed auto update channel picks it up on next launch if you are already on 3.0.x. There is no macOS build; on a Mac, LU Cloud runs in the browser. New here? The handbook takes you from installer to first generation, and the install chapter lists the file names.

The easiest local AI image generator, the Create tab Qwen-Image 2.1 joins
Character LoRA training, locally, where the trigger words in the new tab come from
Image to image with local AI, the lane that now has a maskless option
v2.5.9 release notes, the previous release write-up on this site

Locally Uncensored 3.0.1. AGPL-3.0 licensed. Built by PurpleDoubleD. Bug reports and feature requests on GitHub Issues or in the Discord.