Creative Generation with Local AI in 2026

Chatting with a local LLM is useful. Making things with local AI is addictive. This is the third article in my 2026 series — after the hardware and the language models — and it covers the creative side: images, video, voices, and two things I couldn't have written about a year ago: 3D-printable models and old-photo restoration. Everything here runs on your own machine, free.

ComfyUI is the hub

All roads lead to ComfyUI. It's a node-graph editor — you wire boxes together into a workflow — and every model below plugs into it. The learning curve is real but front-loaded; load a template workflow, generate something, then start rewiring. The images on this blog all come out of it, usually finished with a pass through GIMP.

My prompting workflow hasn't changed, because it works: I ask an LLM — a cloud one like Claude, or my own local Qwen — to write me a detailed positive prompt and a negative prompt for the image I have in my head, then paste both into ComfyUI. A picture is worth a thousand words, but a good prompt is worth a hundred bad generations.

Images in 2026

• FLUX.2 [klein] — the headline of early 2026: a 4B distilled model under a true Apache 2.0 license, runs in about 13GB of VRAM, and handles text-to-image, editing, and multi-reference prompting. Quality that punches far above its size.
• Z-Image Turbo — Alibaba's 6B speedster; a 1024px image in a few seconds on modest hardware. When I'm iterating on an idea, this is the scratchpad.
• Qwen-Image — the 20B one you reach for when the image needs readable text inside it: posters, signs, labels. Nothing else local comes close.
• SDXL — old faithful. Still the deepest library of community styles and LoRAs on Civitai, still happy on smaller cards.

Video

Local video finally crossed from party trick to actually useful. The workhorse is Wan 2.2 — the 14B version, GGUF-quantized, produces real quality 720p on a 16GB card, and there's a 1.3B version for 8GB cards. (Alibaba's newer Wan 2.5 and 2.6 didn't get open weights, which I'll grumble about quietly and move on.) The other one to watch is LTX-2, the first open model that generates synced audio and video in one pass. Use the CPU-offload nodes for the text encoder and you claw back several GB of VRAM.

Voices

Text-to-speech went from robotic to eerie. Chatterbox is the one I point people to first: MIT-licensed, tiny (the Turbo model is 350M parameters), and in Resemble AI's blind tests listeners preferred it to ElevenLabs almost three to one. VibeVoice from Microsoft does long-form — up to 90 minutes with four distinct speakers, basically a podcast generator. IndexTTS2 is the emotive one with precise duration control, though read its license before any commercial use. All three have ComfyUI nodes, and all three do voice cloning from a short sample.

Which is the moment to say it plainly: clone your own voice, or a voice you have permission to use. Full stop.

New trick #1: 3D-printable models

This one merges two of my hobbies. Hunyuan3D 2.1 runs locally in ComfyUI and turns a single image into a 3D mesh. The full pipeline with PBR textures wants 20GB+ of VRAM — but here's the beautiful part for us 3D-printing people: a printer doesn't care about textures. Mesh-only generation runs in as little as 6–10GB. Generate the mesh, export it, clean it up in Blender or Meshmixer, make sure it's watertight, slice, print. An object you described in a sentence, sitting on your desk in plastic by dinnertime. (TRELLIS is the MIT-licensed alternative if you want to compare outputs.)

New trick #2: restoring old photos

The image-editing models — Qwen-Image-Edit and FLUX Kontext — turn out to be quietly wonderful at photo restoration. Scratches, dust, tears, faded contrast, even colorizing black-and-white: you feed in the damaged scan and a plain-English instruction, and out comes the photograph your family remembers. It runs on 8GB of VRAM. I've been scanning boxes of old family prints, and this is the closest thing to time travel my computer does. If you're the designated archivist of your family — and if you're reading this, you probably are — this alone justifies the hardware.

---

A few closing utilities and rules of the road: HandBrake for converting whatever video format the AI hands you, Civitai for models and workflows, and the same thing I said in the original Medium article — with this AI power comes great responsibility. Use it in a reasonable, beneficial manner. Don't be malicious.

Everything in this series — the hardware, the models, these tools — will still be on your machine and still working years from now, whatever happens to subscription prices in the cloud. That's the whole idea.

Enjoy, and happy free AI generation!