A darkroom for your GPU

A stone lighthouse at dusk on a wet pier, generated by Latent with Z-Image Turbo
Z-Image Turbo · 9 steps · seed 1561448469

Write a prompt or drop in a photo, and watch the image develop on your own GPU. Free and open source. Nothing leaves your computer.

Download Latent Version 1.1.0 · Windows, Linux and Apple Silicon Macs

The contact sheet

Latent with Z-Image Turbo loaded: a lighthouse frame on the canvas, the model bar and VRAM meter above it, the prompt and Generate below, and the frame's recipe on the right
The canvasModels and the memory they use along the top, the frame in the middle, its recipe on the right. Z-Image Turbo making a 1024 × 1024 frame on an RTX 5070 Ti.
Latent's Models window listing Z-Image Turbo, FLUX.3 klein, Qwen-Image 2.1 Turbo, FLUX.2 klein 4B and SANA-Sprint 1.6B, each marked Fits this PC
Models that fitLatent reads your GPU and memory and offers each model in a variant that fits, with its download size, VRAM and license before you commit.
Latent running SANA-Sprint 1.6B on the CPU: the top bar counts RAM, and every model shows the RAM it would need
No GPU, still worksSANA-Sprint 1.6B on the CPU, about 37 seconds a frame. The bar counts RAM instead of VRAM, and each model shows what it would need.
Latent in the light theme with a lighthouse frame made by Z-Image Turbo
DaylightThe light theme. On Linux with Caelestia, Latent takes on your shell's live colors instead.
Latent in the light theme running on the CPU
Every frame keeps its recipeModel, prompt, size, seed and steps go into each saved PNG. Press R on any frame to make it again, seed included.

Keyboard commands

  • GenerateCtrlEnterEach frame appears in its tile a step at a time, so a bad seed shows in the first seconds. The model stays loaded for the next one.
  • Cancel the runEscLatent stops between steps, so you're never waiting out a frame you already know is wrong.
  • Use as sourceEAny frame becomes the photo for the next edit. Make, pick, refine, without saving and importing again.
  • Paste a photoCtrlVDrop, paste or import a picture and describe the change. Pick several to combine them. FLUX.3 klein keeps what you didn't ask to change.
  • Reuse the recipeRModel, prompt, size, seed and steps come back exactly as they were, ready to tweak one thing.
  • Save the frameCtrlSFrames wait in a temporary cache until you keep one, as PNG or JPEG. Nothing is saved behind your back.
  • Next frameJKStep through a batch with J and K or the arrow keys, and delete the misses as you go.
  • Share to your phoneCtrl,From Settings, open Latent in a browser on your laptop or phone while this PC's GPU does the work. Home Wi-Fi or Tailscale. Off until you start it.
13 s

for a 1024 × 1024 frame with Z-Image Turbo on an RTX 5070 Ti. SANA-Sprint does it in 7.

Timed on a real PC

2 GB

of VRAM is enough to stream a model through your GPU a block at a time.

0

accounts or telemetry. Prompts and photos stay on your PC.

Will it run on your PC?

Set your graphics card's memory and see how each model would run. Times are seconds per 1024 × 1024 frame, measured with Latent.

  • SANA-Sprint 1.6B

    GPU 7 s · CPU 37 s

    Generate in 2 steps · 5.3 GB VRAM · Apache 2.0

    7 s

  • FLUX.2 klein 4B

    GPU 12 s

    Generate and edit in 4 steps · 9.8 GB VRAM, 6.2 GB as FP8 · Apache 2.0

    12 s

  • Kromoss

    GPU 12 s

    Generate in 8 steps · 13.7 GB VRAM, 9.0 GB as FP8 · Apache 2.0

    12 s

  • Z-Image Turbo

    GPU 13 s

    Generate in 9 steps · 14.2 GB VRAM · Apache 2.0

    13 s

  • FLUX.3 klein

    GPU 17 s

    Generate, or edit with up to 10 photos, in 16 steps · 12.9 GB VRAM · free under US$5M revenue

    17 s

  • Qwen-Image 2.1 Turbo

    GPU 20 s

    Generate and edit, strong on text and layouts, in 8 steps · 10.5 GB VRAM · non-commercial

    20 s

Measured on an RTX 5070 Ti with an NVMe SSD and a Ryzen 7 7800X3D. Streaming from RAM needs the whole model in RAM (12 to 22 GB); streaming from disk needs less, and a hard drive makes it much slower. On Apple Silicon models run whole in shared memory. Each model keeps its own license, which applies to what you make with it.

How a frame develops

A diffusion model starts from pure noise and removes a little of it at every step, until a picture is left. Latent shows you each step in the frame's tile as it happens, the way a print comes up in the tray.

Noise

Every run starts from random noise set by the seed. The same seed and prompt give the same frame.

Steps

Two steps for SANA-Sprint, sixteen for FLUX.3 klein. Watch them land and press Esc on a miss.

Frame

The finished frame waits in the cache with its recipe until you keep it.

Filed under questions

PriceIs it really free?

Yes. Latent is MIT-licensed, with no accounts, credits or limits. The models it downloads keep their own licenses: most are Apache 2.0, FLUX.3 klein is free under US$5M revenue, and Qwen-Image is non-commercial.

HardwareWhat does my PC need?

On a 16 GB card, start with Z-Image Turbo. On 8 GB, FLUX.2 klein 4B or SANA-Sprint. Models too big for your card stream through it in 2 to 5 GB of VRAM, and with no usable GPU they run on the CPU. Plan on 7 to 33 GB of disk per model, plus uv to download them.

MacDoes it run on a Mac?

Yes, on Apple Silicon with macOS 14 or later, from the .dmg. It's newer than the Windows and Linux builds and hasn't been tested as widely yet. Intel Macs aren't supported.

PrivacyDoes anything leave my computer?

Your photos, prompts and images never do. Latent goes online only to check GitHub for updates and to download the models you choose from PyPI, PyTorch and Hugging Face. Sharing to your phone is off until you start it.

GPUsWhich graphics cards work?

NVIDIA, AMD Radeon on Linux, Intel Arc and Apple Silicon, or the CPU. If a driver or runtime is missing, Latent tells you which package to install for your distro.

Real peopleCan I edit photos of other people?

Edit your own photos, or someone else's with their permission. Sexual or degrading images of real people, fake photos meant to deceive, and impersonation are illegal in many places and not what Latent is for. Saved PNGs are marked as AI-generated. See Responsible use.

Version 1.1.0 for Windows, Linux and Apple Silicon Macs. About 4 MB, plus the models you choose.