v2.3.0 — ComfyUI plug & play, image-to-video on 6 GB VRAM, and uncensored image gen in a local AI desktop app

#webdev #ai #opensource #productivity

Today we're releasing v2.3.0 of Locally Uncensored — an open-source desktop app for running AI locally. Chat, code agent, image generation, video generation. One app, no cloud, no Docker.

This update is the biggest one yet. Here's what changed and why.

ComfyUI Plug & Play

If you've ever tried setting up ComfyUI, you know the pain. Custom nodes, workflow JSONs, model paths, Python environments breaking every other week.

We built auto-detection and one-click install directly into the app. Launch it, and if ComfyUI isn't installed, one click handles it. If it is installed, the app finds it automatically. No YAML files, no CLI commands, no config.

The Dynamic Workflow Builder runs 14 different strategies under the hood. It checks which custom nodes you have installed and builds the correct pipeline automatically. You just write a prompt and hit generate.

Image-to-Image and Image-to-Video

v2.3.0 adds Image-to-Image (I2I) for all supported image models — SDXL, FLUX, Z-Image. Upload a source image, set a denoise strength, write what you want changed.

Image-to-Video (I2V) supports FramePack F1, CogVideoX, and SVD. FramePack is the highlight here — it runs on 6 GB VRAM through next-frame prediction. That's consumer GPU territory. Upload an image, get a video. On hardware most people already own.

Model Bundles

Instead of hunting for model files, checkpoint URLs, and compatible LoRAs, we're building one-click download bundles. Image and video models, verified and tested. The bundle list keeps growing — we're adding more over time as we verify them.

VRAM-aware filtering shows you only what your GPU can actually run. Three tabs: Lightweight, Mid-Range, High-End. No more downloading a 12 GB model to find out your 8 GB card can't load it.

Z-Image — Uncensored Image Generation

Z-Image Turbo generates images in 8-15 seconds with zero content filters. No safety classifiers, no prompt rejection, no "I can't generate that" messages. It's a research model and it does what you tell it. Works for both text-to-image and the new image-to-image pipeline.

New LLM Support

We added GLM 5.1 (754B MoE, MIT license — just released this week), Qwen 3.5, and Gemma 4 to the Discover models section. The hardware-aware onboarding now recommends models based on your actual GPU VRAM instead of a generic list.

What Was Already There

For context, the app already includes:

Uncensored AI chat with 20+ provider presets (Ollama, LM Studio, vLLM, KoboldCpp, and 8 more local backends, plus cloud fallbacks)
Codex coding agent that reads your codebase, writes files, runs shell commands — up to 20 tool iterations
13 agent tools — web search, file I/O, shell execution, code execution, screenshots, system info
A/B model compare — same prompt, two models, side by side with parallel streaming
Local benchmarks — one-click tokens/sec measurement with leaderboard
RAG — upload PDFs, DOCX, TXT and chat with them
Voice chat — push-to-talk STT + sentence-level TTS streaming
Memory system, thinking mode, granular permissions, 20+ personas