InfiniGen brings Krea 2, Stable Diffusion XL, Qwen Image and MiniMax H3 into one calm studio for Windows: images, edits and video with sound, rendered on your own NVIDIA GPU.
Nothing renders in the cloud.
One studio for
01Models
Every model arrives with a profile: its files with checksums, tested defaults and ranges, the sizes it was trained for and what it is best at. Switch models and the whole page follows.
Fast, photographic, cinematic.
The official Turbo models and the community's mixes and finetunes, each with settings that were tested on a real GPU. Every LoRA key format ComfyUI accepts, plus a one-click repair for the ones that would quietly do nothing.
The library you already have.
Base and refiner, the popular finetunes, Illustrious and Pony, and 4-step DMD2 for instant drafts. Your SDXL LoRAs come along.
Edit with words.
Qwen Image 2.1 text to image and ten edit cards, from cutouts and outfit swaps to a pose studio and character sheets. Qwen Image Edit 2511 works from up to three pictures.
Stills, and motion with sound.
Text to image, edits and reference edits, and video with stereo sound: from a prompt, from a first and last frame, or from up to nine reference pictures. Plus the H3 Detail Refiner for a final polish.
Bring your own. Point InfiniGen at the ComfyUI models folder you already have; it only reads it and notices new files without a restart. Paste a Hugging Face or Civitai link to download more.
02Video
MiniMax H3 turns a prompt, a picture or a handful of references into video with stereo sound, right on your GPU. Give a character a line to speak, and say how to deliver it.
PromptNorthern lights ripple over a still fjord at night, slow push in, snow on the peaks
Spoken line“Look up. It's starting.” · whispered
03Presets
A preset is a complete recipe: model, LoRAs, a detailed prompt with fill-ins, size and finish, and settings. Pick a place and a light, add your own touch, and run it with one click, or one key.
Fill in
4 × 3840 × 2160 · Balanced · ≈ 50 s each
This demo paints in your browser. In InfiniGen the same preset runs the real model on your GPU.
4K Holiday Themed Wallpaper
Northern Lights over a Frozen Lake
Minimal Desert Dunes
Deep Space Nebula
City Skyline at Blue Hour
Rainy Neon City Night
Halloween Moonlit Pumpkin Patch
Abstract Fluid Glass
Plus Cozy Cabin Window, Autumn Forest Light, Starry Mountain Night for phones, and your own: save any image or any page of settings as a preset.
04Gallery
Projects keep your work in real folders on your disk. Each image carries its prompt, model, LoRAs and seed, so one key brings any of it back: the same look with a new seed, or the very same image again.
05Details
The small things that keep you making pictures instead of managing them.
Models, presets, prompts, projects and every command, a few keys away. Single-letter keys act on the image in front of you, and the cheat sheet is one ? away.
Kohya, diffusers, LoKr and DoRA keys all work. Files saved with PEFT keys get a one-click repair instead of silently doing nothing. Names, trigger words, strengths and previews are yours to edit.
Enhance a prompt or Describe an image with any OpenAI-compatible model, like Ollama or LM Studio on your own network.
Standard, 2K or 4K, and Fast, Balanced or Best. The final pixels and the time are on the button before you click.
Pause, cancel and retry. Close the window and InfiniGen keeps rendering from the tray; what's queued carries on after a restart.
Pick one: this page follows.
06Privacy
InfiniGen runs its own ComfyUI engine on your GPU. The only things that leave your machine are the downloads and lookups you ask for, and the prompt assistant you choose.
Images, gallery, settings and logs live in one data folder. Back it up, move it, keep it.
API keys are encrypted with Windows DPAPI and only ever sent to their own sites.
Civitai lookups stay off until you turn them on, and character LoRAs are never sent at all.
The local API answers InfiniGen's own window. Requests from other web pages are refused.
07Get started
One installer, no admin rights and nothing to install first. Updating is running a newer installer; your engine and your data stay put.
It brings its own Python, adds the Visual C++ runtime only if Windows lacks it, and leaves your PATH and registry alone.
PyTorch, ComfyUI and pinned custom nodes, about 3 GB, set up in a window that shows its progress. Ten to thirty minutes.
Krea 2 Turbo downloads on the first start, or point InfiniGen at the model folder you already have. Then press CtrlEnter.
08Questions
No. Every image and video is rendered by the engine on your own GPU and saved in InfiniGen's data folder. The only network traffic is what you ask for: model downloads, Hugging Face and Civitai lookups, and the prompt assistant if you connect one.
Yes. Choose your ComfyUI models folder, or any folder with checkpoints, diffusion_models and loras inside, in the installer or later. InfiniGen only reads it; its own downloads go to its data folder. Files without a profile get one automatically from their headers.
An NVIDIA card with a driver from the 580 series or later. Krea 2 Turbo wants 16 GB of video memory or more. InfiniGen was built and tested on an RTX 4090 with 24 GB.
Not today. InfiniGen is a Windows app built around NVIDIA's CUDA.
InfiniGen keeps running in the notification area and queued images keep rendering. Click the tray icon to bring it back, or make the X exit instead in Settings.
Yes. Import a ComfyUI workflow and it shows up as a model, with its inputs mapped to the Generate page.
InfiniGen for Windows is almost ready. Every model, every frame, on your own GPU.