01
Overview
Meet the local image and video generator, its web UI and the creative workflow.
Home / Documentation
Learn your local studio.
A clear path from your first setup to the engine underneath. Find a guide, check a reference or explore how UnstableField works.
Try: hardware · first launch · model packs · API
New to UnstableField?
Start here. Make something next.
Follow this path before your first model download.
02
Check your hardware
Find your RAM / VRAM profile and understand the evidence behind supported tasks.
03
Install and run
Prepare uv and node, run the installer and complete first setup.
Use the studio
A guide for every part of making.
End-user guides for the things you can do today.
Create
From a prompt to an image or video. Choose Image or Video, Fast / Balanced / Best and an aspect ratio.
Chat
Keep the creative thread with streamed replies, image input, prompt rewriting and the generate tool.
Models
Understand the catalogue and memory tiers. Install packs, import compatible local files and manage LoRAs.
What can I do?
The short version.
Useful answers before you dive into the guides.
- Generate images with Z-Image Turbo, FLUX.2 klein or FLUX.1 schnell (token required). Create text-to-video or image-to-video with Wan2.2 5B and A14B Lightning, subject to hardware eligibility.
- Yes. Chat includes threads, streamed replies, image input, prompt rewriting and a generate tool. Its own text runtime shares the GPU with generation jobs.
- Models supports compatible local-file imports, LoRAs and text-model installs. Importing a model by Hub repository ID is not supported. The extended catalogue lives behind Show restricted models.
- Generated images and videos live in the local Library. Generation runs on your machine; model downloads and first-chat runtime libraries need network access. Logs and metrics stay local.
Model licenses vary. Open weights are not a promise of unrestricted commercial use. Check each model’s license and requirements before using its outputs.
Build and understand
Under the hood.
Contributor and developer references. Separate from the end-user setup path, for when you want to understand or extend the engine.
Engine & architecture
How the local torch + diffusers engine fits together, without ComfyUI.
API reference
Integrate with the local service using its documented request and event contracts.
Calibration & parity
Understand how plans are checked and how measurement evidence is collected.
Know the boundaries
Evidence, not assumptions.
See what has been measured, what is estimated and where the project is heading.
Read the RTX 4070 Ti results.
Calibrated device runs and capped-memory measurements are distinct. Estimated profiles are not real laptop benchmarks.
What’s next is not what’s shipped.
Follow project direction separately from current capabilities. HunyuanVideo 1.5 and LTX-2.5 runners remain unmeasured.
View on GitHub