Field notes.
Dispatches from building Vision Studio - a free, open-source desktop app for local AI image and video generation. Local-first workflows, open tooling, the engineering behind each release, and a clear-eyed case against the cloud subscription treadmill. No marketing fluff. Just what runs, and why it runs better on your own GPU.
The index.
Every dispatch, filed by capability and timestamped. Tap a card to read the full note.
Garbled signs and melted lettering are not a drawing problem. They start in the text encoder, which often never sees the letters of a word at all. What the research shows, and how to pick a model for an image that has to say something.
The EU AI Act transparency rules applied from 2 August 2026, California's AI Transparency Act the same day, and every major platform now asks for an AI label. Who has to do what, and why a local generator puts the job in your hands.
Re-rolling a whole image to fix one hand throws away everything that worked. Inpainting regenerates only the masked part, and one number, strength, decides how far it moves. What that number means, step by step.
Running the model on your own machine does not make you the author. The US Copyright Office, a federal appeals court and the Supreme Court have now drawn the line between prompting and authorship. Where it sits, and how to stay on the right side of it.
AI upscalers and face restorers do not find pixels the camera missed. They paint plausible ones from a learned prior, and their own papers say so. When that is exactly what you want, and when it is the wrong tool.
A LoRA file is a small set of low-rank changes to a model you already have. What that means for weights, stacking, base-model matching and licences, with the numbers from the paper that introduced it.
Published speed charts say torch.compile and int8 quantization make diffusion models fly. On a real desktop, with changing sizes and a 12 GB card, the free wins are precision and attention. Here is the measured order of returns.
Peer-reviewed measurements put AI image generation at 2.9 kWh per thousand images on data-centre GPUs. What that means on your desk, what it costs at 18 cents a kWh, and how to measure it yourself.
Research papers quote video generated faster than real time on an H100. Your desktop is not an H100. The three open video models worth running locally, what each one is for, and the VRAM each one needs.
A prompt is a weak way to say where things go. ControlNet and IP-Adapter turned layout and reference into inputs the model has to follow, and the U.S. Copyright Office says human input you can see in the output is what counts.
Almost half of professionals admit pasting non-public company data into generative AI tools. "Runs locally" is the answer many vendors give. Here are five checks that tell you whether it is true, run against our own app.
Cranking steps to 100 and guidance to 15 feels like buying quality. The research says modern samplers converge in 15 to 25 steps, and high guidance trades away variety. Here is how to find the point where more stops helping.
Share a seed and expect the same image? PyTorch itself says identical seeds are not guaranteed to reproduce across releases, platforms, or CPU and GPU. Here is what a reproducible render actually has to record.
Open weights are not one licence. FLUX.1 [dev] lets you sell its outputs but not run it for revenue. SD 3.5 is free until $1M in revenue. And prompts alone do not make an image yours. Read this before client work.
JFrog found around 100 malicious models on Hugging Face, one opening a reverse shell on load. Pickle-format weights can run code the moment you open them. What safe loading looks like, and how Vision Studio enforces it.
For local image and video generation, VRAM decides what runs at all; speed only decides how fast. September 2026 Steam data shows 16 GB narrowly ahead of 8 GB. Here is what each tier can run.
One commit, three build machines, two working apps. The unpinned dependency that broke our macOS release, and the YAML parser hiding inside the update path of the shipped binary.
Creators are stacking Midjourney, Runway, Pika, and Firefly into a four-figure annual bill, paying every month to rent models that run on a GPU they already own. Here is the real math, and the open-source way out.