Kinoloom 0.1.1

The installer now says what failed and why. It uses the model files already on your PC and checks your graphics card. The movie side now renders through the Video Studio. You can write the story and its video prompts with the built-in AI writer or with any chat AI. Cuts open on frames made from your cast's pictures, and the QA checks that the cast still looks like itself. The full list is in CHANGELOG.md.

Install

• Failures are reported. When a step fails, the window says which step, what happened and what to do, with the log lines behind it. The report is saved in logs\install\LAST-FAILURE.txt, and logs\install\install.log records every run. INSTALL-MAIN.bat modes [1] and [3] end with NOT FINISHED (exit code 2) when a part did not install, instead of DONE.
• Model files you already have are used. Mode [3] finds the weights already on your PC, under any name and in any folder, such as a ComfyUI models folder. It links them into place (no extra disk space), uses them where they are, or copies them only if you agree. Your files are never renamed, moved or changed. Settings › Model files in the web app does the same.
• Your graphics card is checked. The install's last check names your card and driver and what they can't run. The launcher doesn't start a ComfyUI server your card can't run. The model lists mark such a model as "not on this graphics card", with the reason.
• Qwen-Image 2.1 gets the build your card runs. On a GTX 16- or RTX 20-series card or newer with driver R580 or newer, it runs on PyTorch's CUDA 13.0 build, which supports RTX 50-series cards. On a GTX 900- or 10-series card, or with a driver from R560 up to R580, it stays on the CUDA 12.6 build that 0.1.0 installed.
• start.bat doesn't start an unfinished install: it shows the last failure report and offers to finish it. Folder paths with %, ^, &, !, ' or non-ANSI letters work.

Movie

• Movie shots render the way Video Studio renders them. On LTX-2.5 and MiniMax H3, each shot or take gets the request the Video Studio page would send: the cast as named references, a start image or the previous clip's last frame, and prompts written for that render.
• Write with any AI. The story and the video prompts can come from the AI writer (any OpenAI-compatible server: a cloud provider, or Ollama or LM Studio on this PC), or from any chat AI by copy and paste. Both ways use the same prompts.
• Production renders the prompts it has. If only some shots have prompts (say, the first batch pasted from a chat AI), production renders those and stops where they end. On the Shots page, Use a chat AI for the next batch copies the brief for the next prompts; paste the reply there and press Continue production to go on from there.
• The story is written in takes. A take is one continuous camera passage, rendered as one video: up to 6 shots on LTX-2.5 and 4 on H3, or a length you set per project. A cut without a start image of its own opens on a frame Qwen-Image 2.1 fast makes from your cast's pictures, when its prompt includes an opening image (the AI writer writes one).
• The cast stays itself. The QA compares the faces in each LTX-2.5 and H3 shot with the cast's pictures and sends a shot that lost its cast to review.
• Movie and Studio sides. The sidebar switches between Movie (Casting, Cast & World, Story, Shots, Render, Review, Film) and Studio (Video Studio, Image Studio, Library). When production stops for you, a dialog, the tab title and a desktop notification say so.

Security

• hydra-core 1.3.8 in the web app's environment, for GHSA-wf6j-7hvr-gr93. SAM-2 needs it; no Kinoloom code uses Hydra itself. Running INSTALL.bat after the update installs it.

Coming from 0.1.0

Your projects, settings, model files and Python environments stay where they are.

1. Get the new files: git pull in the Kinoloom folder, or unpack the 0.1.1 ZIP and copy its contents over your Kinoloom folder, replacing the files there.
2. Run INSTALL.bat again. It checks every part and rebuilds only what changed. From 0.1.0 that is the Qwen-Image 2.1 environment, which moves to the CUDA 13.0 build on a GTX 16- or RTX 20-series card or newer with driver R580 or newer: about 2.6 GB to download, roughly 5 to 30 minutes depending on the connection. An older card or driver keeps the environment it has. The web app's environment gets one more pip run, so this run needs the network.
3. Start Kinoloom with start.bat.

Two things to know:

• Your tool\settings.yaml is kept as it is. Settings added in 0.1.1 use their defaults (the values in tool\settings.example.yaml). In 0.1.0 the AI writer pointed at a local Ollama (http://127.0.0.1:11434/v1, llama3.1:70b), and your settings still say so. If you don't run Ollama, open AI writer on the Story page and pick your provider, or write by copy and paste.
• Graphics cards. LTX-2.5 and MiniMax H3 need a GTX 16- or RTX 20-series card or newer and driver R580 or newer, as in 0.1.0. The render farm's models (the SDXL image models, Z-Image, post-processing) don't run on RTX 50-series cards. Their CUDA 12.6 build has no kernels for those cards, which was already true in 0.1.0; 0.1.1 says so instead of failing at render time.

Tested

On one PC (Windows 11 Pro 25H2, RTX 3070 8 GB, driver 616.64), from a source archive of a release candidate (the release differs from it in hydra-core 1.3.8, the Shots page's next-batch button, one disk figure in bootstrap.ps1 and two lines that matter only on Linux):

• A first install, in a folder named Kino & 100% ^ !'s (7) é, ended in 19 min 58 s and downloaded 6.3 GB; the folder then held 16.2 GB. Its last check named the card and driver and the parts they run.
• Mode [3] linked LTX-2.5 and Qwen-Image 2.1 (13 files, 44.3 GB) in from a folder that held them, in 5 s, without copying them.
• Renders: that install's own ComfyUI servers rendered a Qwen-Image 2.1 still and an LTX-2.5 clip from it.
• Downloads blocked: mode [3] ended with NOT FINISHED (exit code 2) instead of DONE: after 31 s when the proxy's name did not resolve, after 3 min 13 s when every connection was refused. With the downloads open again, the refused part finished.
• An older card: an install as on a GTX 1080 with driver 560.94 (through a stand-in nvidia-smi) built Qwen-Image 2.1 on CUDA 12.6, which rendered a still (on the RTX 3070), and the web app's video model list gave the card as the reason LTX-2.5 and MiniMax H3 can't run.
• Coming from 0.1.0: 0.1.0 with a project in it, the 0.1.1 files copied over it, then INSTALL.bat: done in 9 min 47 s, with Qwen-Image 2.1 moved to CUDA 13.0 and the settings kept; the project opened and saved with 0.1.1. As on a GTX 1080, the same update (without a project) kept Qwen-Image 2.1 on CUDA 12.6 and took 2 min 38 s.
• Tests: 1176 unit and integration tests pass on this release (1 skipped: it needs Windows Developer Mode for symbolic links).

A second PC, with a GTX 1080, installed a release candidate and started it; nothing was rendered there.

CHANGELOG.md has the details, and what was not tested.

Known limitations

• The graphics card check has run on two cards: an RTX 3070, and a GTX 1080 that installed and started a release candidate without a render. Other cards and drivers are covered by tests and by an install with a stand-in nvidia-smi, not by a render.
• Seamless chains are experimental. A whole movie has not rendered through the new movie path on the GPU, and no H3 movie shot has.

See Known limitations in the README for the full list.

Kinoloom 0.1.0

The first release of Kinoloom, a local AI film studio for one Windows PC with an NVIDIA card. It's an alpha: it runs on the machine it was built on (RTX 3070 8 GB, 64 GB of RAM), and it hasn't been tested on other hardware yet.

What you need

• 64-bit Windows 11. Windows 10 version 1803 or later is untested.
• An NVIDIA card with 8 GB or more, with driver R580 or newer for LTX-2.5 and MiniMax H3 (R560 or newer for the rest).
• 64 GB of RAM. The big video models keep part of their weights in system RAM.
• About 17 GB of disk for the install, plus the models you choose. LTX-2.5 alone is a 28.8 GB download.
• An internet connection. You don't need Python or Git: the installer brings its own.

Install

1. Under Assets below, download Source code (zip).
2. Extract it: right-click the ZIP and choose Extract All. Don't run anything from inside the ZIP.
3. Move the folder that has INSTALL.bat in it somewhere with a short path, like C:\Kinoloom.
4. Double-click INSTALL.bat and press a key. It downloads about 9 GB and takes roughly 15 to 45 minutes. Windows may ask for administrator rights once or twice: to install the Microsoft Visual C++ runtime if your PC doesn't have it, and to turn on long paths if the folder's path is long. If the install stops, run it again; finished steps are skipped.
5. Get the model weights: run install\INSTALL-MAIN.bat, choose mode [3], press Enter to keep the studio folder, then pick the models. Enter alone installs the three recommended ones: LTX-2.5, Qwen-Image 2.1 and post-processing. Before it downloads anything, it shows the plan, the free disk space and each model's licence, and asks you to confirm.
6. Double-click start.bat. The web app opens in your browser at http://127.0.0.1:8080.

How long things take

On the RTX 3070 8 GB, once the model is loaded:

• a 5-second 1080p clip with LTX-2.5: about 6 minutes
• a 16-second take in 4 links with LTX-2.5: about 28 minutes
• a Qwen-Image 2.1 still: 38 to 78 seconds

MiniMax H3

Kinoloom supports MiniMax H3 (all 11 tiers, cast references and long takes), but the installer never downloads its weights. Its licence grants no use in the EU, the UK, South Korea or the US. If you're outside those regions, see MiniMax H3: bring your own weights in the README.

Known limits

• Windows and NVIDIA only.
• Tested on one PC. If you run it on other hardware, a report helps a lot: see Help test more hardware.
• Long takes (seamless chains) are still experimental.
• It's slow. These are big models on a small card.

The code is MIT, with the exceptions listed in THIRD_PARTY_NOTICES.md. Every model keeps its own licence.

More: the README, SHOWCASE.md for every timing and QA result, and CHANGELOG.md.