LumaBrowser bundles a local image-generation server built on stable-diffusion.cpp. Text-to-image and image editing run entirely on your own machine, no cloud service, no per-image fee, and nothing leaves your device. Describe a picture and get a render; open one you already have and repaint just the part you want.
This is local inference, not a hosted API. The generation server runs on your own machine. Your prompts and the images you create, including any photos you bring in to edit, never leave the device. There is no inference provider in the loop, no upload of your source images, and no per-image cost.
The same local server covers both ends of the workflow: making an image from nothing, and reworking one you already have.
Describe an image in plain language and get a render. Pick a generation model, set the size, and the local server produces the picture on your GPU, no queue, no credits, no upload.
Open an image and change it without starting over:
Models download on demand from Hugging Face; nothing is bundled until you choose it. The catalog spans fast turbo models, high-fidelity renderers, and a dedicated editing model, each listed with a short note and an approximate download size.
| Model | What it's for | Approx size |
|---|---|---|
| Z-Image Turbo | Fast 8-step bilingual text-to-image - a quick default for everyday renders. | ~6 GB |
| Qwen-Image-Edit 2509 | 20B editing model for character / subject continuity via reference images. | ~18 GB |
| Stable Diffusion 1.5 | The classic - runs on 4 GB cards, native 512×512. | ~2 GB |
| SDXL 1.0 | Higher fidelity, native 1024×1024. | ~6.5 GB |
| FLUX.1-schnell | Photorealistic 4-step generation. | ~12 GB |
An optional Anima anime model add-on is available too. And the catalog isn't a ceiling: you can bring your own single-file .safetensors or .gguf model, or import a Hugging Face repo directly. Approximate sizes are for the on-demand download; the exact figure depends on the variant you choose.
LumaBrowser keeps a generation model and an edit model loaded at the same time, with separate VRAM reservations. You can create and refine in the same session without paying the reload tax between steps.
Your text-to-image model stays loaded, Z-Image Turbo, SDXL, FLUX.1-schnell, or whatever you picked, ready to render the next prompt instantly.
An editing model like Qwen-Image-Edit holds its own VRAM reservation in parallel, so jumping from a fresh render to an inpaint or reference-image edit doesn't unload anything.
Image generation isn't a separate tool you have to switch into. From the built-in AI Chat, the agent has generate_image and edit_image tools; just ask in the conversation.
Describe what you want in chat and the agent calls the local generation model. The result comes back as an image artifact you can open full-size or download, rendered on your machine, not fetched from a service.
Point the agent at an image and ask for an edit, image-to-image, an inpaint, or a reference-guided change. It runs the edit model and returns the new image as an artifact in the conversation.
For setting up the chat model itself, hardware-aware picks, resumable downloads, streaming chat, see Local AI.
The image server uses the same hardware backends as the rest of LumaBrowser's local AI, so it runs across the major vendors and Apple Silicon, with VRAM-aware placement so you can pin image models to specific GPUs.
The fast path on NVIDIA cards; the larger models in the catalog are happiest here.
Vulkan covers AMD, Intel, and other GPUs so you're not locked to one vendor.
M-series Macs run the same models locally; no external GPU required.
Bigger models want more VRAM, the editing and FLUX models in particular. VAE tiling helps Qwen-Image-Edit fit under 24 GB. VRAM-aware placement lets you pin a specific image model to a specific GPU, which matters when you keep both the generation and edit slots resident. The same hardware detection that powers Local AI applies here.
Free, local, and private, text-to-image and full image editing on your own GPU, with a curated catalog and your own models welcome. No cloud service, no per-image fee, nothing leaving your device. Pair it with Local AI and the built-in APIs, or head home to see everything LumaBrowser does.