v1.1.0

What's new

Video generation. Quartermaster renders video now, on the same stable-diffusion.cpp backend that already does images: write a prompt, or hand it a starting image, and get a clip back. MiniMax-H3 and Wan 2.x are both served, each with the components and generation defaults they expect wired up for them.

A clip takes minutes, not seconds, so a render is a job rather than a request. The model is pinned in VRAM for as long as its render is running, nothing evicts it halfway through, and closing the tab does not kill the work: the clip finishes and is waiting when you come back. Length is picked in seconds, LoRAs come from the model's own folder (neither family is distilled, so a turbo LoRA plus a lower step count is the single biggest lever on how long a render takes), and a size or length that probably will not fit warns you in orange before you spend the time, with the arithmetic behind the guess.

The ceiling on a single clip is VRAM, and there is a way around it: Continue from here takes a finished clip's last frame and makes it the next one's first. Each clip costs the same no matter how long the finished video gets, so a long sequence is a chain of short renders. Models trained for it also take an end frame and travel toward it.

The 3D tab. TRELLIS.2 arrived last release as a backend and an API endpoint. Now it has somewhere to live: drop in an image, get a textured mesh, and orbit it in the tab without downloading anything or leaving for another program.

A model browser that stays out of the way. Browse no longer defaults to Trendy, which had been quietly hiding every established model behind "what came out recently". The file picker lists everything in a repo instead of GGUFs alone, so a model whose VAE, text encoder or projector ships as safetensors can be assembled without leaving for a web browser. Video repos are found, LoRA repos are no longer collapsed into their parent, and an interrupted download survives a restart: a half-finished transfer comes back paused with its real progress instead of disappearing.

Backends you can pin. The backend picker on a model now lists every build you have installed, not one row per engine, so a model can be pinned to a specific sd-server or llama.cpp build while the rest of your models follow whatever the class default is. And a pin is now honored everywhere: embedding, speech recognition and hand-declared image models used to quietly ignore it and follow the default anyway.

Plus the usual smaller things: a model's playground button knows what kind of model it is (a video model offered "Chat" before), the playground rail fits its labels without reflowing the tab, and editing a message uses the full width of the bubble.

One thing to expect after upgrading: the config is regenerated once (genVersion v70 to v74). Context length, offload and tensor splits for a model you have not touched can come out differently than before, and for models sharing a box with the new classes that is the point.

Changes

• docs: a Video generation article (#53)
• playground: edit at full bubble width, wrap long code
• playground: a rail that fits its labels and never reflows the tab
• docs: the 3D tab's subsystem note
• playground: a 3D tab, image in and an orbitable mesh out
• models: one table decides a model's playground button
• playground: anchor the settings panel to the composer bottom
• video: frame refs move into the sent turn, not stay in the composer
• video: emit --temporal-tiling only where the VAE implements it
• video: send frame refs as data URLs, matching sd-server's own client
• video: move the frame pickers into the composer
• video: first/last frame conditioning, and clip chaining
• video: add temporal-tiling and stream-layers knobs, on by default
• video: warn in orange on settings that probably will not fit
• video: pick clip length in seconds, and open up the ladders
• video: stop the UI size cap from silently clamping a model's launch size
• hub: find safetensors video repos, and stop collapsing LoRA repos
• video: add the LoRA picker to the Video tab
• video: correct the MiniMax-H3 generation defaults
• hub: document MarkSelection and the Trendy default
• hub: list every repo file, and stop defaulting Browse to Trendy
• hub: keep video-understanding VLMs out of the Video tab
• video: serve video generation as its own model class
• autogen: honor per-model backend pins in every class
• backends: list every installed build in the model backend picker
• hub: restore interrupted downloads on startup
• autogen: let a declared text-encoder pin win the width tie

Full changelog: v1.0.7...v1.1.0

Instructions

Download the setup program for your platform:

• Windows -- quartermaster-setup-windows-amd64-v1.1.0.exe
• Linux -- quartermaster-setup-linux-amd64-v1.1.0 (or -linux-arm64-v1.1.0)
• macOS, Apple silicon -- quartermaster-setup-darwin-arm64-v1.1.0

The unix ones need chmod +x first. There is no signed macOS build yet, so
clear the quarantine flag before the first launch:

xattr -d com.apple.quarantine ./quartermaster-setup-darwin-arm64-v1.1.0

SHA256SUMS covers every file here.

The other files, and who they are for

quartermaster-windows-amd64.exe and quartermaster-linux-amd64 /
-linux-arm64 / -darwin-arm64 are the bare server binary, not an
installer. Nothing about them creates a Start Menu entry, a desktop shortcut or
an uninstall record. They are published for two reasons:

• the in-app updater downloads one to replace the running binary, so a release
  without them silently stops every existing install from updating;
• a headless Linux or macOS box can run one directly: chmod +x, point it at
  a models folder, and install backends from Settings.

Installing for the first time? Take a quartermaster-setup-... file from
the instructions above.