v3.0.7: Your Characters Everywhere, ComfyUI Workflows as Saved, and Training on the Mac

Locally Uncensored 3.0.7 came out on October 19, 2026. A character you trained now works in every part of Create and in chat, and a Mac can train one too. ComfyUI workflows import in the format ComfyUI saves, LU installs the node packs they need, and every setting of an imported workflow can be changed in Create. Edit takes up to ten images. A chat can be duplicated or branched from any message.

Many of the requests in this release came from the Discord and from GitHub issues. Our check in the real app ran on one Windows machine, and the Mac parts ran with the real engine on one Mac. What has only run in automated tests is listed at the end of this post.

The short version:

• A trained character works in Image, Edit, every video tab and the chat tools. A Mac trains its own, a cloud character runs locally, frames from a video become a training set, Edit can use the training photos, and a character can change clothes. Low VRAM training is new for cards under 12 GB
• ComfyUI workflows import as ComfyUI saves them, subgraphs included. LU installs and updates node packs, and every node setting of a workflow is editable in Create
• Edit with Qwen-Image 2.1 takes up to ten images
• Duplicate a chat, branch from a message, and keep earlier answers as versions
• On a Mac: training, Talking Character, Motion Control, Music, Enhance Image, Remove Background, masks in Edit, HiRes fix and more for video
• Character cards, a ChatGPT import, LoRAs from a link, Check files and Repair for models, several OpenAI-compatible backends side by side, and a round of fixes for group chats and personas

Your character in every part of Create and in chat

When Create runs on ComfyUI, Image, Edit and every video tab have a character chip next to the model chip. On Image and Edit the character's LoRA and trigger word go into the run, and the pick sets the Z Image model the character was trained for. Where no fitting model is installed, the line under the character says so with a button to switch, and Create does not start a run that would leave the character out (issue 161). The same line asks you to describe the key features in the prompt too, because the trigger word alone often does not hold the look. Edit with a mask does not take a character.

On Video, Animate, Extend, Multishot, MiniMax H3, Motion Control and Talking Character, LU first makes a picture of the character and then starts the clip from it, shown as Step 1 of 2 and Step 2 of 2. A start image of your own wins. A video model that only takes text, or a tab with no video model installed, says so before anything runs. In chat, the image and video tools take a character by name or trigger word.

More ways to make and use a character:

• Frames from a video. The lightbox and the large view of a video offer Use frames for a character. LU takes 8 frames spread evenly over the clip, 4 to 20 with the slider. Untick the frames that do not show the figure, then train with the rest in Character Studio or save them as a character
• A cloud character on your machine. Use locally on the LU Cloud character shelf saves a Z Image character into the LoRA folder of this machine, so the local Use character shelf lists it. If a file of that name is already there, LU stops and says so instead of replacing it. Characters of other model families are refused with the reason
• Edit with the training photos. The reference strip in Edit has a Trained characters row. A click adds up to three of the photos a local character was trained on, at most 1024 pixels each. A character trained before 3.0.7 has no photos there
• A character that changes clothes. The Train tab (Windows, Linux and Mac) has a short text field under each photo. Name there what changes from photo to photo, like clothes, hair or place. It goes into that photo's caption after the trigger word, so the character keeps its face and body but not one outfit. Describe photos fills the empty fields with the local chat model picked in Chat, when that model can see images. It never sends the photos to a cloud model
• Low VRAM training. On Windows and Linux a second setting sits next to the training length. Normal is the 12 GB recipe as before. Low VRAM trains at 512 px instead of 768 px, keeps 28 of the model's 30 blocks and the saved activations in system memory, and trains a smaller LoRA (rank 16 instead of 32). Each step takes clearly longer and fine detail comes out softer. A card with less than 11 GB is pointed at Low VRAM, which lets cards from about 8 GB train
• Delete a character. The Use character shelf has a delete button for a character trained on this machine. It asks first and moves the file to the Recycle Bin or Trash

Character Studio also names only the base files that are missing, with their size, and downloads only those. Before, it showed about 19 GB whatever was missing.

ComfyUI workflows in the format ComfyUI saves

The Workflows view now takes a workflow saved with Save or Export in ComfyUI, not only one saved with Save (API Format). LU reads the node settings from the running ComfyUI, so ComfyUI has to be started for this, and the import says so when it is not. Reroute and Primitive nodes are resolved, notes are left out, bypassed nodes pass their input through and muted nodes are left out. Subgraphs are unpacked the way ComfyUI unpacks them when it runs a workflow, nested ones and ones used more than once included. That means the stock ComfyUI templates built on subgraphs, all Z-Image templates among them, can be imported. Workflows with the older group nodes, nodes ComfyUI does not know, or settings that do not fit the installed node version are turned down with the reason. Choose files picks several at once, and Choose folder imports every .json in a folder, each file with its own result (Discussion 124).

Node packs, managed from LU

Settings, AI Backends has a new section, ComfyUI node packs, on Windows and Linux. It lists every pack in the custom_nodes folder with where it came from, its branch, commit and date, and marks packs that are switched off. A pack installs from a link to its repository on github.com, gitlab.com or codeberg.org (https only). LU first tells you that a pack is code that runs inside ComfyUI, then clones it and installs its requirements.txt into the Python of your ComfyUI, with the log on screen. Update pulls the newest version and installs the requirements again. A pack LU pins to a tested version stays on it. Remove moves the folder to the trash, so nothing is deleted for good, and Restart ComfyUI loads the change.

When an imported workflow uses nodes ComfyUI does not know, the Workflows view offers the pack for every node whose pack LU knows, and names the others with the way to the link field. A ComfyUI on another machine is not managed from here (issue 14, Discussion 72).

Every setting of a workflow, in Create

Under Advanced settings, a workflow from the Workflows view now shows a list called Workflow parameters: each node with its values, such as the VAE, CLIP skip, LoRA strength, denoise or any value of a custom node. Numbers keep the limits ComfyUI states, and lists offer the choices ComfyUI has. What LU sets on its own (prompt, seed, size, steps, CFG, sampler, scheduler, model) is named per node and not offered twice. The VAE and text encoder stay on Auto until you pick one. Changes are kept per workflow and go into every run, and Reset goes back to the values in the file. Nodes fold, and a search finds a setting by name. Built-in workflows are not affected. This is on Windows and Linux, since a Mac runs no ComfyUI workflows (issue 8).

Your own workflow in Motion Control

Motion Control always used LU's own graph, even with a workflow picked for its model. A workflow imported in the Workflows window and paired with a Wan Animate model now runs there. LU fills in the model, prompt, seed, character picture and driving video and leaves everything else as saved. The import reads the model from a GGUF loader and the clip from Load Video or Video Helper Suite's Load Video, files a Wan Animate graph under that model family, and the Workflow picker finds it for a Motion Control model. On Video, Animate and Motion Control the workflow button now reads Use your own ComfyUI workflow.

Edit with up to ten images

With Qwen-Image 2.1 on your own machine, the reference strip under the source now takes up to nine more images, up from three. On Windows and Linux this runs through ComfyUI, on a Mac through the built-in engine. The tiles wrap into rows, and the hint names them as a range, for example "image 2 to image 10". From four extra images on it asks you to say what each one brings, for example "the scarf from image 3".

With more than four images in all, every image is scaled down so that together they take no more memory than four did. At a 1024 canvas, ten images run at 640 pixels each. On ComfyUI the result then comes out at that smaller size too, and the hint says so. On a Mac the result keeps the size you set, and the strip offers only as many images as the Mac's memory takes at that size. Images past that stay in the strip and are marked not sent. On a Mac, ten images at 640 pixels each ran with the real engine in 78 seconds.

Qwen-Image-Edit 2509 and 2511 stay at three images in all, LU Cloud at five photos, and a trained character still brings three of its photos.

Duplicate, branch and keep earlier answers

The chat list and its right click menu have Duplicate chat: a full copy that opens at once, with its persona, system prompt, group members, sampling values, documents and the Agent and Code settings. Every message has Branch from here, which starts a new chat with the history up to that message. Regenerate, and Save & resend on an edited message, keep the earlier answer as a version, and the arrows on the answer switch between them (issue 174). Duplicate chat is not offered in Remote Access.

Editing an earlier message no longer sends it or deletes the rest (issue 172). Before, the check mark resent the message at once and deleted every later message for good. The edit field now has Cancel, Save and Save & resend. Save changes the text only and keeps every later turn. Enter saves, and Ctrl or Cmd with Enter resends.

Create on the Mac (Beta)

The Mac build makes images and video without ComfyUI. 3.0.7 adds the tools below. All of them are in beta.

• Training. Character Studio's Train tab trains a Z Image character from 4 to 100 photos with the Mac's image engine and puts it into the Mac's LoRA folder, where Use character, the character chip and chat find it. The first run downloads a training adapter of 340 MB. Training is slow on a Mac: on an M5 Pro one step takes about 7 seconds, so the Standard setting (1200 steps) runs a little over two hours. Use frames for a character hands its photos to the same training
• Characters. Use character lists the characters in the Mac's LoRA folder. A pick brings the Z Image model the character was trained for, and Image, the video tabs and the chat tools take the character as they do on Windows. A video starts from a picture of the character made on the Mac
• Talking Character. A portrait and a voice recording become a clip whose lips follow the voice, as long as the recording and up to 10 seconds, with LTX-2.3 4-bit. It needs one more download of 11.3 GB, offered on the tab, and a Mac with 24 GB or more. It is slow: a 4 second clip took about 10 minutes on an M5 Pro
• Motion Control. A motion clip and a start image become a video that follows the motion, with LTX-2.3 4-bit and the IC-LoRA Union Control (about 650 MB, offered on the tab). LU turns the motion clip into edges first. A 2 second clip at 512 x 512 took 103 seconds on an M5 Pro
• Music with ACE-Step 1.5, once the music engine (about 10 GB) is installed from the Music rail of the Model Manager. A run takes the style, the lyrics, a length from 10 seconds to 4 minutes and the seed, and puts the track in the gallery
• Enhance Image with SeedVR2 3B (7.28 GB) from the Model Manager, 2x or 4x, with the longest side held to 4096 pixels. A result larger than the memory takes is refused: at 4x a 1024 picture needs about 22 GB, and a smaller Mac is told to pick 2x or a smaller picture
• Remove Background with the subject detection built into macOS, so it needs no model and no Python. A picture without a clear subject gets a sentence instead of an empty image
• Edit with a mask. With Qwen-Image 2.1, Edit offers Paint mask, Edit strength and Mask edge feather, and a result offers Edit with mask. On an M5 Pro with 48 GB, a hard mask at full strength changed only the masked area in 189 seconds
• HiRes fix as a second pass at the larger size with a model that edits, such as Qwen-Image 2.1. The latent upscale method is not offered on a Mac
• Video. Where the model's entry allows it, the LTX and Wan video models take a stack of LoRAs and a width and height, and Wan takes 1 to 60 steps. Extend Video is a local tab while a video model that takes a start image is installed, and the Shots slider reaches the LTX-2.5 packs. The large LTX models now render at the size Create shows instead of 512 x 512
• A Qwen prompt enhancer for Qwen-Image 2.1: two packs of about 6 GB each, one for new pictures and one for edits, rewrite the prompt before the run. The edit enhancer also looks at the picture

Also on the Mac: switching from Cloud to Local keeps Talking Character and Character Studio open, Music no longer fills the memory while it decodes, Extend Video shows the Mac's own video settings, and in the computer control permissions the Reset and ask again link shows whenever a permission is missing.

Mask edit, HiRes fix, Remove Background, Enhance Image, an image with a LoRA, both prompt enhancers, LTX-2.3 video at a chosen size, Extend Video, Music, training, Talking Character, Motion Control and an edit with ten pictures ran with the real engine on one Mac with 48 GB. They have not been clicked through in a built app yet.

The Mac we tested on is an M5 Pro with 48 GB on macOS 26.3.1. Expect long runs there: the Standard training setting takes a little over two hours, Talking Character needs about 150 seconds per second of clip, and a 2 second Motion Control clip at 512 x 512 takes 103 seconds. In the app, computer control (not fully tested yet on macOS), Voice, Create in the chat and new cloud models also carry the beta label.

Chat

• Sampling values per persona and per group member (issue 175). The persona editor takes its own temperature, top P, top K, min P, max tokens, repeat, presence and frequency penalty and DRY multiplier. An empty field follows the chat, and values set in the chat's Sampling popup still win. The popup lists each speaker with the values that hold for it and marks a value its backend does not read
• The persona picked in Plugins is the one that is sent (issue 170). A chat now remembers its persona, and chat, group chat, Agent and Voice all send what the menu shows
• Group chats keep track of who is in them and who spoke (issue 165, issue 169). A member whose model is no longer listed sits out and shows as unavailable with a remove button. A server that lists the same model twice gives one row, and a turn keeps the persona it was written as
• Character cards become personas. A card as PNG or JSON (versions 1, 2 and 3) imports as a persona with the card picture as its tile. A new chat with it opens with the card's greeting, and arrows switch between alternate greetings until you write. A card persona exports back as a version 2 JSON file
• A ChatGPT export becomes real chats. Settings, General, Import from other chatbots takes the .zip of ChatGPT's data export or the conversations.json inside it. Each conversation you pick becomes its own chat with its title, date and the thread you saw. Images and files are not copied, and a note marks where they were. A second import adds nothing twice, and a chat you went on with in LU is never replaced
• An attached text file fits the model's context (issue 171). A text file was cut to its first 4000 characters without a word. The chat now keeps up to about a million characters and gives the model as much as fits next to the conversation, and tells it how much it got. A new setting, Attached file limit, caps that share
• Any file can go into the chat. A file that is not text says above the chat that models cannot read it and points to Agent or Code. There the file goes into the working folder, and the model gets its name, type, size, SHA-256 and path, with the instruction not to run it unless you ask. The limit is 256 MB per file, up from 64 MB
• A chat compacted more than once keeps one summary (issue 151). A new compaction merges the earlier summary and the turns since then into one current state, and puts back requests, open tasks, paths and numbers the merge dropped
• Reasoning effort for local models (issue 176). Low, Medium, High and Max now reach the LU Engine, LM Studio 0.4.8 and later, KoboldCpp and a llama.cpp server of your own, each in the form it reads. Elsewhere the button is greyed and says why
• Think budget next to Think in the sampling popup and in Settings, Generation. It works with the LU Engine and with Claude models. For Ollama, LM Studio, other OpenAI-compatible servers and LU Cloud the field is off and says so
• Chat Reply Timeout (issue 177). Settings, General sets how long a chat waits for the first word, up to 24 hours, or Off. The default stays 600 seconds, and Stop always ends the wait
• Read aloud speaks what the answer shows (issue 160). Read aloud and the automatic reading take the same cleaned text the bubble shows, so reasoning markers are no longer spoken
• A picture made in the chat shows on the desktop. A current ComfyUI refuses the app window's direct load of a picture, so a chat image showed as a broken frame. The chat now fetches it through the app, as Create already did, also when the model repeats a ComfyUI picture in its answer
• Sending while you dictate ends the recording first
• The User-Agent of web search can be set under Settings, Search Provider
• Remote Access can carry your own name for the browser tab on the phone (issue 19), a paired phone can show the name of this computer, and the Remote Access texts point to the right buttons

Create

• Enhance Image runs locally on Windows and Linux, on ComfyUI, 2x or 4x. Without an upscale model the stage offers Real-ESRGAN x4plus, a 67 MB download that LU checks against its checksum
• Gallery folders. A folder is only a label: the files stay where they are, and removing a folder keeps its pictures. While a folder is open, new results land in it. Under a result, Prompt and settings lists the negative prompt, model, steps, CFG, sampler, seed and size, with one button that copies all of it
• Images per run sit next to the Create button: 1 to 8 locally, 1 to 4 in LU Cloud
• Animate with a reference picture on Wan 2.1 VACE. Phantom, MAGREF, the VACE add-on modules and Wan 2.2 VACE Fun are listed with the reason they do not run yet
• Each local video family has its own clip length. MiniMax H3 goes up to 362 frames, 15 seconds at 24 frames per second, and the slider moves in the steps the runs use
• A model too big for this machine says so on its tile in the Model Manager, with one Try LU Cloud button. Create shows a line of the same kind, and it can be closed. Nothing of this shows in Cloud mode or with Show Cloud features in Local mode off
• Create honours the image and video timeouts from Settings, and asks ComfyUI whether a run is still going before it stops one
• Create sees model files you copied in by hand when you come back to the app
• The mask editor shows Ctrl+Z and Ctrl+Y on Windows and Linux, and the settings of a music track name its seed
• Smaller fixes: a wide result fits the stage with its prompt below it, Cloud mode no longer shows the load error of the local models, a tooltip no longer opens when the focus comes back to its button, Rewritten by names the chat model the way the picker does, the first image of a list keeps its file name, a second Enhance of the same picture no longer says Queued when ComfyUI takes it from its cache, Erase and Edit show the mask state as it is, the model chip of Character says to pick a character first, and the voice maker row of the lip sync tool stays inside its panel

Models and backends

• A LoRA from a link. Models, LoRAs, Get new takes a CivitAI model link or a Hugging Face link to a .safetensors file. LU reads the name, base model and size before anything downloads, checks the SHA256 when the site states one, and puts the file into the LoRA folder. Checkpoints and other models that are not LoRAs are refused, and a file of the same name is never replaced
• Check files and Repair (issue 113). Image and video bundles and chat models of the LU Engine have a Check files button that compares every file with the catalog by size and, where listed, SHA256. Repair downloads only the broken or missing files again. A wrong file goes to the Trash or Recycle Bin first
• Several OpenAI-compatible backends at the same time (issue 173). Adding KoboldCpp while LM Studio or the LU Engine is on no longer pushes the other one out. Each backend gets its own row with its own name, address, optional key and model IDs, and a group chat can mix models from several of them
• Model IDs a server serves but does not list can be added on the provider card (issue 112)
• A GGUF file from anywhere on your drive runs as a chat model on the LU Engine. LU links it into the LU Engine folder without copying, so it has to be on the same drive
• The voice engine and LU's own Python can live on another drive. Settings, Model Storage has a row for them. Move to this folder copies the files, checks the copy, starts it once and only then puts the old files in the Recycle Bin or Trash
• Context sizes from the server. A model whose server names no context length is no longer held at 4K (issue 166). LM Studio keeps the context size you picked for a model (issue 168). Japanese, Chinese and Korean text is counted closer to its real token cost, and the context counter of a group chat names the speaker it counts for (issue 167)
• GPU layers -1 lets the LU Engine fit the layers itself on AMD and Intel cards, against the free memory it measures (issue 156). On older cards that answer in garbage, LU first turns Flash Attention off and moves to the processor only when that does not help (issue 156)
• ComfyUI starts when no Python on your machine has PyTorch, for example with Python 3.15 as the only one. Install and Repair environment then build ComfyUI's environment from LU's own Python 3.12, which needs no administrator rights
• A render no longer starts LM Studio when it is closed, and a chat model with image input no longer warns at every render
• Motion Control and Talking Character models show under Installed. Wan Animate, Wan VACE and Wan S2V files are listed under Video, and the row says which Create tab runs them. Keep the original file name: a renamed file goes through the plain video graph, which cannot run it
• A search in Get new covers both catalogue tabs while you type, so FramePack shows up from Mainstream too
• Importing a model follows the LU Engine folder you set under Model Storage
• A download whose partial file was already complete no longer fails on every Retry with HTTP 416 (issue 113)
• Voice and dictation may use the microphone on Linux (issue 162)
• Settings, Privacy names the real data folders on Windows and Linux

Security

ComfyUI and LM Studio, when LU starts them, answer to LU only, so other pages open in your browser can no longer read them. Each time LU starts the LU Engine, the engine gets a key that only this run of LU holds, so other pages can no longer run its model or read what it is doing. A page the agent fetches reaches the model marked as data from that page, not as instructions to follow.

What we have not tested

The changelog has the full list. These are the points to know before you rely on something:

• Our check in the real app ran on one Windows machine with an NVIDIA card. For characters it used a stand-in LoRA, so how well a real character keeps its face has not been checked, and step 2 of a character video has not run with a real video model
• In the app on Windows we checked the workflow import, Workflow parameters, node packs, a LoRA from a link, the ChatGPT import, the timeouts, images per run, the tab name for the phone and a GGUF file from the drive. Subgraphs, Update of a node pack, a real ChatGPT export and Edit with ten images on Windows and Linux have run in automated tests only
• Duplicate chat, Branch from here, the kept answers, Save of an edit, sampling values per persona, the persona fix, the group chat fixes and attached text files have run in automated tests only. So have Check files and Repair, the text per photo and Low VRAM for training. Low VRAM has not been measured on a card with 8 GB
• Use frames for a character, the Trained characters row in Edit with a real training set, and Use locally against LU Cloud have run in automated tests only
• Your own workflow in Motion Control was checked with tests on QuantStack's Wan 2.2 Animate example workflow, not yet in a real run against ComfyUI. The chat picture fix was checked with tests, not yet on a Windows machine with ComfyUI Desktop
• The clip lengths per video family and Wan 2.1 VACE have not run with a real video model
• Reasoning effort for local models, the Chat Reply Timeout, several backends side by side and the line about LU Cloud for models too big have run in automated tests only. The think budget ran against a real LU Engine; with Claude models it has run in automated tests only
• The microphone fix for Linux has not been tried on a real machine, and GPU layers -1 has not been checked on an AMD or Intel card. Character cards were checked with Holo and Qwen models, not with Gemma
• Training on a Mac has run with 10 photos and 100 steps on one Mac with 48 GB. The Standard and Thorough settings, Macs with less memory and stopping a training in the middle have not. The large LTX models (64 GB and more) have not run on a real Mac, nor has anything on a Mac with less than 48 GB. Video LoRAs and multishot have not run with real models on a Mac
• The Windows installers of this release are not code signed, so SmartScreen and some antivirus engines can warn about them. Is Locally Uncensored safe? explains the warning and how to check a download.
• The update in the app from 3.0.6 to 3.0.7 was not checked through on any system before the release

Download

Each link loads the file directly.

• Windows: Locally.Uncensored_3.0.7_x64-setup.exe
• Mac with Apple Silicon, macOS 14 or newer: LU_3.0.7_aarch64.dmg
• Linux: AppImage, deb or rpm

Already on 3.0.x on Windows or Linux? The signed update channel offers new releases on start and waits for your click before it restarts. We have not checked the step from 3.0.6 to 3.0.7 in the app, so if it does not arrive, take the installer above. New here? The handbook takes you from installer to first generation, and the install chapter explains the SmartScreen warning and how updates arrive.

Related

• Train a character LoRA locally, the guide to Character Studio
• ComfyUI beginners guide, if workflows are new to you
• What a Mac can run locally in 2026, by memory tier
• Locally Uncensored for Mac, the current state of the Mac build
• v3.0.6 release notes, the previous release with Voice and computer control

Locally Uncensored 3.0.7. AGPL-3.0 licensed. Built by PurpleDoubleD. Bug reports and feature requests on GitHub Issues or in the Discord.

In the handbook

• Create, every tab and what it needs
• Character Studio, training a character and using it
• Install and update, the SmartScreen warning and how updates arrive
• Settings and troubleshooting, every settings tab in one line each and where the log file is