The Local AI Toolkit: A Guide by Use Case

In order to help our customers get a head-start in Local AI we've compiled a running list of well-regarded local and self-hosted AI software!

This list is organized by "Use-cases" of what you're actually trying to do, but due to the constantly evolving nature of Local AI technology some of the entries point directly to software and models and others are more like suggestions and keywords that will help you find what you're looking for.

Most of the image, video, and 3D tools below run through ComfyUI, which acts as a shared engine — adding a new capability to ComfyUI is usually just a matter of loading a different model rather than installing new software. Text and chat work generally runs on Ollama, and document work on AnythingLLM.

AI Video Creation

Video is the most GPU-hungry corner of local AI, but 2026's open-weight models have closed most of the gap with closed-source tools like Sora or Veo. Between text-to-video generation, frame interpolation, and AI-assisted upscaling, a genuinely local video pipeline is now achievable — no per-render fees, no watermarks, and nothing uploaded anywhere. The tools below cover everything from raw generation to polishing the final cut.

Text-to-video generation — ComfyUI + Wan 2.2/2.7 (GitHub). Alibaba's Wan is currently the strongest open-weight text-to-video model, with a mixture-of-experts architecture that handles complex, multi-step prompts well.
Image-to-video workflows — Wan 2.2 I2V (same repo above) or LTX-2.3. LTX is notably lighter on VRAM and faster to iterate with than Wan, at some cost to fidelity.
AI animation — HunyuanVideo, run via ComfyUI. Tuned for human motion and expressive animation, making it a strong choice for character-driven clips.
Video enhancement — REAL Video Enhancer. Free and cross-platform, bundling upscaling, denoising, and frame interpolation into a single app.
Frame interpolation — REAL Video Enhancer (RIFE engine) or Flowframes (Windows only). Both convert lower frame-rate footage to smooth 60fps+ output locally on a GPU.
Video upscaling — REAL Video Enhancer or Video2X. Both wrap Real-ESRGAN for video, upscaling frame by frame and reassembling with FFmpeg.
AI-assisted filmmaking — HunyuanVideo for shot generation combined with REAL Video Enhancer for post-processing; ComfyUI is typically used to chain these into a repeatable pipeline.
Automated video production — A ComfyUI pipeline chaining Wan/LTX generation into REAL Video Enhancer for upscaling. This is less a single app and more a workflow built from the pieces above, which is exactly what ComfyUI's node graph is designed for.

Advanced Image Creation

For anyone pushing past casual image generation into serious creative or commercial work, the local image-generation stack has matured enormously. Complex ComfyUI workflows, high-resolution output, and models with genuinely permissive commercial licences now put professional-grade artwork within reach of a single well-specced GPU.

Professional AI artwork generation — ComfyUI + FLUX.2 (Black Forest Labs). FLUX remains the quality benchmark among local image models for prompt adherence and detail.
Large image models — Qwen-Image 2.0 (Hugging Face). Alibaba's 7B model generates natively at 2K and is particularly strong at rendering legible text within an image, an area where FLUX struggles.
Complex ComfyUI workflows — ComfyUI-Manager, the standard tool for installing and managing custom nodes so multi-model workflows don't break on missing dependencies.
High-resolution creative projects — FLUX.2 or SDXL with a hi-res fix pass in ComfyUI, followed by Upscayl for a final resolution boost.
Commercial content creation — FLUX.2 [klein] (Apache 2.0), the first FLUX variant that's genuinely commercial-clean with no revenue cap, unlike FLUX.1 [dev], which requires a paid license for business use.

Advanced AI & Development

Running larger language models locally isn't just about privacy — it's also the foundation for building custom AI applications, coding assistants, and research tools that don't depend on a subscription or an internet connection. This section covers the software that turns a downloaded model into something genuinely useful to build on.

Larger local language models — Ollama, the standard tool for pulling and running large open models (Llama, Qwen, DeepSeek, etc.) via a single command and an OpenAI-compatible API.
AI coding assistants — Cline, a VS Code extension that reads a project's files, plans changes, and runs terminal commands. Works fully offline when paired with a local Ollama model.
Advanced document analysis — AnythingLLM. Documents are dropped in and indexed locally, so they can be queried privately with no data leaving the machine.
Multi-agent AI workflows — n8n (self-hosted) for visual workflow orchestration, or CrewAI for code-first multi-agent setups. Both pair with local Ollama models.
AI application development — LM Studio as a local model server, paired with LangFlow for visually building application logic on top of it.
Model experimentation — LM Studio, whose GUI makes it easy to download and compare different models and quantisations side by side.
AI research and learning — Open WebUI, a ChatGPT-style front end for Ollama with retrieval-augmented generation and web search built in.

3D & Digital Creation

Generating usable 3D assets locally, rather than modelling everything by hand, is one of the newer and more exciting local AI capabilities. From a single reference image to a textured, game-ready mesh in seconds, these tools are reshaping early-stage 3D and character work.

Advanced 3D asset generation — Hunyuan3D 2.1. Generates full PBR-textured meshes from a single image or text prompt, fully open-sourced.
Game asset creation — Hunyuan3D or TRELLIS (Microsoft). TRELLIS handles trickier topology — holes, thin surfaces — better than Hunyuan3D, which suits irregular game props.
AI-assisted animation — Cascadeur, a physics-aware 3D animation tool with a free tier that uses AI to auto-generate believable in-between poses and secondary motion.
Product visualisation — Hunyuan3D for image-to-3D generation, with the resulting mesh brought into Blender for lighting and rendering.
Virtual environments — TRELLIS or Hunyuan3D for individual assets, assembled manually in Blender. There isn't yet a strong dedicated local model for generating a whole environment at once, so this remains an asset-by-asset process.
Digital character creation — TRELLIS for the base mesh, with Stable Diffusion or FLUX used beforehand for texture and concept art.

Audio & Media Production

Voice cloning, music generation, and audio restoration have all quietly become genuinely production-ready on consumer hardware. Whether cleaning up a noisy recording or generating a full backing track, the tools in this category run entirely offline and, in several cases, now outperform their well-known commercial rivals in blind testing.

AI voice generation — Chatterbox (Resemble AI, MIT licence). Offers zero-shot voice cloning from roughly five seconds of reference audio, and has outperformed ElevenLabs in blind listening tests.
Podcast enhancement — Resemble Enhance, which denoises and upsamples speech recordings ahead of publishing.
Speech cleanup — DeepFilterNet. Very low latency and CPU-friendly, and better at avoiding the "underwater" artifacting that older noise-removal tools produce.
Audio restoration — Resemble Enhance's super-resolution mode, which reconstructs missing high-frequency detail — useful for reviving old or heavily compressed recordings.
AI music experimentation — ACE-Step 1.5, currently the leading local music generation model. Runs under 12GB VRAM and supports LoRA fine-tuning on a custom dataset of songs.
Sound effect creation — Stable Audio Open (Stability AI), purpose-built for foley and textural sound design rather than full songs.

Future-Focused Projects

Some of the most interesting local AI use cases don't have a single dedicated app yet — they're built by combining several tools into a pipeline. Virtual influencers, automated content channels, and custom AI-powered tools all fall into this more experimental territory, where the "software" is really an assembled workflow rather than an off-the-shelf product.

Virtual influencers — A combination of ComfyUI (a consistent character via a trained LoRA), Wan/LTX for video, and Chatterbox for voice. No single app currently covers this end to end, so it's typically built as a personal pipeline from these pieces.
AI-powered content channels — The same generation stack above, with n8n handling scheduling and publishing automation between steps.
Custom AI tools — Ollama's local API as the model backend, with n8n or LangFlow as the application layer built on top.
Automation systems — n8n, self-hosted workflow automation with native AI-agent nodes and far more flexibility than typical no-code platforms.
Experimental AI projects — LM Studio or Ollama's model library, generally the lowest-friction way to try a newly released model.

AI Image Creation

From concept art and character design to marketing imagery and logo concepts, this is the everyday, practical end of local image generation. The right model choice usually comes down to licensing and each model's specific strengths — some excel at photorealism, others at rendering clean text or following detailed prompts.

AI artwork generation — ComfyUI + SDXL/FLUX.
Concept art creation — SDXL with ControlNet in ComfyUI, allowing a rough composition or pose sketch to guide the model's rendering.
Character design — SDXL/FLUX with a trained LoRA, letting a specific character's appearance stay consistent across multiple generations.
Product mock-ups — FLUX.2 [klein], chosen for its clean commercial licence.
Logo concepts — Qwen-Image 2.0, currently the strongest local model for rendering legible text and typography within an image.
Marketing imagery — FLUX.2, for the same commercial-licence reasons as product mock-ups.
Interior design visualisation — SDXL/FLUX with ControlNet (depth or canny), so a photo of a real room can guide a generated redesign.
Fantasy and creative artwork — SDXL with LoRAs, a common combination for stylised or genre-specific artwork.
Texture generation for games and 3D projects — Hunyuan3D-Paint for PBR textures on generated meshes, or plain Stable Diffusion for flat/tileable textures.

Photography & Image Editing

Long before generative AI, photo editing already had a strong tradition of specialised, single-purpose tools, and that's carried over into the local AI space. Upscaling, restoration, background removal, and colourisation each have a dedicated, mature open-source tool rather than one do-everything app.

AI photo enhancement — Upscayl, a free, open-source, Real-ESRGAN-based tool comparable to paid options like Topaz for most everyday photos.
Image upscaling — Upscayl, the clear free/local alternative to Topaz Gigapixel.
Old photo restoration — GFPGAN, which specialises in restoring and sharpening faces in old or damaged photographs.
Noise reduction — Upscayl's denoise models, or Real-ESRGAN directly for a lighter pass.
Background removal — rembg (with the BiRefNet model option for the best edge quality on hair and fine detail). Runs entirely offline via a command line or Docker container.
Image editing assistance — Qwen-Image Edit, part of the Qwen-Image 2.0 release, run via ComfyUI. Built specifically for prompt-based edits and inpainting rather than full regeneration.
Colourisation of historic photos — DeOldify, still the most widely used open-source colouriser, though it is an older project with limited active development.

3D Creation & Printing

For hobbyist and small-scale 3D printing, AI-generated meshes are a genuinely useful shortcut from idea to printable object — though the cleanup and print-prep stage still leans on traditional modelling tools rather than AI. This section covers both halves of that workflow.

Convert images into 3D models — Hunyuan3D 2.1 or TRELLIS, both of which take a single photo and output a textured mesh.
Create printable models — Hunyuan3D's mesh output, exported as OBJ/STL and cleaned up in Blender for print-readiness (wall thickness, manifold geometry).
Generate 3D concepts — The same Hunyuan3D/TRELLIS pipeline, useful for testing a shape idea quickly before committing time to hand-modelling.
AI-assisted modelling workflows — TRELLIS 2, specifically for cases where it handles open surfaces and irregular topology better than Hunyuan3D.
Mesh cleanup and optimisation — Largely traditional tooling rather than AI: Blender's remesh and decimate tools applied to raw Hunyuan3D/TRELLIS output.
Create custom figurines — Hunyuan3D-Paint, which produces a textured, print-ready single object.
Prototype product ideas — Hunyuan3D or TRELLIS image-to-3D generation ahead of formal CAD work.
Generate assets for 3D printing — The same pipeline, refined in Blender and then a slicer.

Content Creation

Not every image-generation task needs ComfyUI's full node-based complexity. For quick turnaround work like thumbnails, social graphics, and reference images, a simplified front end can get from prompt to finished image in a fraction of the time.

Create social media graphics — Fooocus, a simplified SDXL front end built for fast, good-looking output without ComfyUI's node graph.
Generate YouTube thumbnails — Fooocus, optionally with a face-consistent LoRA so the same subject appears across a channel's thumbnails.
Develop visual concepts — ComfyUI + SDXL/FLUX, for cases needing more control than Fooocus offers.
Create reference images for projects — Fooocus, for quick single-image generation without workflow setup overhead.

Personal AI Assistant

A local chat assistant is the most accessible entry point into local AI, and it can do far more than casual conversation. Document summarisation, private research, writing help, and translation are all achievable entirely offline once a model and interface are set up.

Run local AI chat assistants — Ollama + Open WebUI. Ollama runs the model; Open WebUI provides a proper chat interface.
Ask questions and brainstorm ideas — The same Ollama/Open WebUI setup, typically with a general-purpose model such as Qwen3 or Llama.
Summarise documents and articles — AnythingLLM. Files are dropped in and summarised without leaving the machine.
Research topics privately — AnythingLLM's built-in web-search agent tool, or Open WebUI's web-search plugin.
Create outlines, reports and presentations — Ollama/Open WebUI for drafting content. There isn't yet a strong local tool for generating the finished .pptx/.docx file itself, so this typically pairs an LLM draft with a conventional office application.
Improve writing and proofreading — Ollama or LM Studio with a writing-oriented model. Quality of prose feedback varies more between models than general chat ability does, so it's worth comparing a couple.
Translate text — LibreTranslate, a dedicated self-hosted translation engine, or a general local chat model for more nuanced, context-aware translation.
Create personalised AI workflows — n8n, connecting local Ollama models into repeatable automations.

Work & Productivity

Local AI is particularly well suited to workplace tasks involving sensitive information, since transcripts, notes, and documents never leave the machine they're processed on. This section covers the tools for turning that privacy advantage into a genuinely useful daily workflow.

AI-assisted emails and correspondence — Ollama + Open WebUI, for drafting locally before pasting into an email client.
Meeting transcription and summaries — whisper.cpp, via a GUI such as Buzz, with the transcript then summarised by a local LLM.
AI-powered note organisation — Obsidian with the Smart Connections plugin, adding local semantic search and AI chat across a notes vault.
Search through your own documents — AnythingLLM, built specifically for private document search and Q&A.
Build a personal knowledge base — AnythingLLM for document-heavy collections, or Obsidian + Smart Connections for a notes-based knowledge base.
Create custom AI assistants for specific tasks — AnythingLLM's workspace/agent system, or an Ollama Modelfile for a lightweight custom system-prompt "persona."

Learning & Development

A local model makes a surprisingly good study companion — patient, available offline, and free to interrogate at length without worrying about API costs. These tools cover everything from a coding tutor to a sandbox for comparing how different models handle the same prompt.

Learn programming with AI assistance — Cline + Ollama, an offline coding tutor that can explain and edit code in place.
Explain complex topics — Open WebUI chat with a strong reasoning model, VRAM permitting.
Practice languages — Ollama chat with a multilingual model, alongside LibreTranslate for quick reference translations.
Explore prompt engineering — LM Studio, whose side-by-side model comparison view makes it easy to see how a prompt behaves across different models.
Experiment with AI tools and models — LM Studio or Ollama's model libraries, the lowest-friction way to try something new.

Creative & Technical

This final category rounds up the more hands-on, technical end of local AI: coding assistants, basic automation, and voice-controlled smart home integrations that keep every command on the local network rather than sending it to a cloud assistant.

Coding assistance — Cline + Ollama, as above.
Generate and explain code — Cline paired with a coding-specific model such as Qwen3-Coder or Codestral.
Basic automation projects — n8n.
Home AI integrations — Home Assistant, the open-source hub for connecting local AI to physical devices, fully self-hosted.
Smart home AI assistants — Home Assistant's Assist pipeline (Whisper for speech-to-text, Piper for text-to-speech, Ollama as the conversation agent) — a fully local voice assistant with no cloud account required, documented here.

General notes

ComfyUI underlies most of the image, video, and 3D entries above — once installed, adding a new capability is usually a matter of loading a different model checkpoint rather than a new application.
VRAM is the main practical constraint across image/video/3D work. FLUX and Wan are comfortable at 16GB+; quantised (FP8/GGUF) versions reduce that requirement at some cost to quality.
Licensing varies more than expected. FLUX.1 [dev] and full-size Wan models are free to download but restricted for commercial use, while FLUX.2 [klein], SDXL, and Qwen-Image are clean for commercial work.