Local AI media ยท plain language ยท your own computer

Plainbrush

Say what you want in plain words and get the picture, the edit or the clip back. Plainbrush drives the open-source image and video models on your own graphics card, with no node graphs to learn and no cloud bill.

> put a vintage armchair by the window in this photo and match the evening light
Started 17 Sep 2026Licence: MIT core (planned, at public repository opening)Runs on 12 GB GPU
30-second demo: two real generations recorded from a browser against a local ComfyUI (the wait is sped up).
14features behind one text box: create, add, remove, relight, extend, restore, animate, music, lip-sync
17tested workflow templates chosen automatically from the request
818automated tests, all passing (17 Nov)
241commits since 17 Sep 2026
0Python packages to install for the core: one standard-library server file
The problem

The best free AI tools are locked behind node graphs

ComfyUI runs the strongest open image and video models on ordinary graphics cards, for free and privately. But it asks the user to wire graphs of nodes, pick checkpoints, and know which model, sampler and add-on fits which job. Most photographers, small studios and hobbyists stop there, and pay monthly for cloud tools instead, sending their photos to someone else's servers.

Plainbrush closes that gap

It reads the request, works out the kind of job (picture, edit, video, music), checks what the user's own ComfyUI has installed, picks and fills the right tested workflow, runs it, and checks the result before showing it. A manual template picker remains for experts.

Watch it work

A real screen recording, start to finish

Driving the real app and a real ComfyUI on an RTX 3060 (12 GB). Long waits are sped up and labelled with the real measured time; nothing here is staged or re-typed.

Real results

One sentence in, finished edit out

Every result below came from Plainbrush on a single RTX 3060 (12 GB), using only the words shown, with no masks, layers or node graphs. They are unretouched. We also ran relighting, style transfer, pasted objects and whole-scene animation; those outputs did not meet our bar yet, so they are not shown.

remove the coffee table
Living room before and after removing the coffee table
BeforeAfter ยท about 3.5 min
replace the floor lamp with a potted plant
Floor lamp replaced by a potted plant
BeforeAfter ยท about 3.5 min
add a golden retriever lying on the rug
Golden retriever added on the rug
BeforeAfter ยท about 3.5 min
put a wooden bench on the studio floor in the middle
Wooden bench added to an empty studio with a reflection on the floor
BeforeAfter ยท about 4 min
extend this photo to the left, more lawn
Patio photo extended to the left with new lawn
BeforeAfter ยท about 4.5 min
a foggy cemetery at dawn, wide angle photo
Generated foggy cemetery at dawn
From words onlyabout 50 s
the same woman sitting in a sunny cafe, candid photo
Reference portrait and the same woman in a cafe
Reference photoNew scene, same person ยท about 3 min
the same man hiking on a mountain trail in a blue jacket
Reference portrait and the same man hiking
Reference photoNew scene, same person ยท about 3.3 min
she smiles softly and turns her head slightly toward the camera
Photo to video (routed to a gentle camera push-in)5 s clip ยท about 10 min

Reference portraits are generated demo characters (no real people). Times are measured on the development machine.

More results

Pictures from words

Generated from the sentence alone, on the same 12 GB card. All characters are fictional.

a lighthouse on a rocky coast at sunrise
Lighthouse at sunrise
From wordsabout 35 s
photo of a rainy night street market, a vendor serving ramen under a neon sign
Rainy night ramen stall
From wordsabout 35 s
a red fox in fresh snow, photographic
Red fox in snow
From wordsabout 80 s
studio portrait photo of a fictional young woman with long red hair and freckles
Portrait of a fictional red-haired woman
From wordsabout 1 min
studio portrait photo of a fictional middle-aged man with short grey hair
Portrait of a fictional grey-haired man
From wordsabout 80 s
the two of them sitting together on a park bench, warm afternoon light
The two fictional characters on a park bench
Two consistent peopleabout 2.5 min
More results

Guided by a photo you already have

The left image is the input; the right is the result. Depth guidance keeps the room or garden layout and changes everything else; pose guidance keeps the body position; style guidance borrows the look of one photo for a new subject.

a futuristic white laboratory with glowing blue screens
Studio turned into a laboratory
Layout fromNew room, same layout ยท about 1 min
a snowy winter garden patio at dusk with warm string lights
Patio turned into a snowy winter garden
Layout fromSame patio in winter ยท about 1 min
a cozy old library interior at night, warm lamplight
Living room turned into a lamplit library
Layout fromSame room, new mood ยท about 1 min
a cheerful chef in a white uniform and hat in a bright kitchen
Chef with the reference pose
Pose fromSame pose, new person ยท about 70 s
a friendly astronaut in a white suit, studio portrait
Astronaut with the reference pose
Pose fromSame pose, new person ยท about 45 s
a quiet harbour town at dusk
Harbour town in the style of the gas station photo
Style fromNew subject, same look ยท about 1 min
a busy open-air market square with colourful stalls
Market square in the style of the patio photo
Style fromNew subject, same look ยท about 40 s
a cozy cottage with a red door in a green meadow, tree beside it
Cottage generated from a sketch
From a rough sketchabout 55 s
More results

More edits, video and sound

add a potted plant in the corner by the window
Plant added by the window
BeforeAfter ยท about 2.5 min
add a wooden stool in the middle of the floor
Stool added to the studio
BeforeAfter ยท about 2.5 min
add a red car parked on the left side of the road
Red car added to the street
BeforeAfter ยท about 2.5 min
the camera slowly moves forward down the street
Photo to video5 s clip ยท about 10 min
the camera slowly pans to the right across the room
Photo to video5 s clip ยท about 9 min
the camera slowly moves forward toward the gas station
Photo to video5 s clip ยท about 10 min
calm ambient piano with soft pads, instrumental
Music from words20 s ยท about 1 min

Held back because they did not meet the bar yet: whole-scene animation (it invents people and animals), text-to-video clips (too little motion or not what was asked), relighting and style transfer on some photos, and pasted objects that float or show seams.

How it works

From a sentence to a checked result

UnderstandA small local language model reads the request; a rule filter backs it up (37 of 40 on the regression set).
Check the machinePlainbrush reads what the user's ComfyUI has installed and what the graphics card and memory can run.
Pick the workflowIt chooses the tested template that fits, fills in sizes, steps and models, and predicts the wait.
Run locallyThe job runs on the user's own ComfyUI with live progress. Nothing is uploaded anywhere.
Check the resultQuality gates reject blank, black, clipped or wrongly sized images before they are shown, and the verdict travels with the result.
Keep itHistory, downloads, variations and a personal asset library, with timings learned from each job.
The product today

What works, measured on an RTX 3060

Plainbrush on a desktop browser after generating a lighthouse imagePlainbrush on a phone-sized screen
FeatureStatusTypical time
Pictures from wordsreadyunder a minute
Add your own object to a photo, scaled, grounded and light-matchedreadyabout a minute
Remove or replace things in a photoreadyunder 2 minutes
Cut out a backgroundreadyseconds
Add something described in words ยท relight a photo ยท extend a photo ยท restore or upscalebetaseconds to 2 minutes
Music clips (instrumental, 20 s)betaunder a minute
Camera move across a photo (video)beta6 to 10 minutes per 5 s
Lip-sync, talking photo, style/pose/depth guidance, consistent facesexperimental2 to 20+ minutes

Status labels are the project's own test verdicts on real photos; experimental features stay switched off until the user asks for them.

Why it is different

Built to be trusted

Local and private

Photos stay on the user's machine unless they choose to connect an outside AI service. No accounts, no per-image fees, no upload of personal pictures.

Hardware-aware

Every feature is gated by what the user's graphics card and memory can actually run, and the page predicts the wait from the user's own past jobs.

Quality gates

Each finished image is checked automatically (blank, black, clipped, wrong size) before it is shown, and the verdict travels with the result.

Licence-honest

Every template carries a licence tier. A commercial-only mode refuses anything with non-commercial terms, so businesses can use it safely.

No bundled assets

Plainbrush ships no model weights and no third-party assets. Users link their own libraries read-only; nothing is used for training.

Tested like a product

818 automated tests (all passing on 17 Nov 2026), a feature contract with ask-or-assume rules, and benchmarks against the best public community workflows.

Features

What it does now, and what comes next

Available now

  • Pictures from words, fast or quality mode, aspect presets, variations
  • Remove, replace or add things in a photo by describing them
  • Paste your own object into a photo, scaled, grounded and light-matched
  • Extend a photo beyond its edges
  • Cut out a background
  • Restore and upscale old or small photos
  • Relight a photo from a named direction (beta)
  • The same person or two people across new scenes
  • Style, pose, depth and sketch guidance
  • Camera moves across a photo as video; scene and face animation (beta)
  • Instrumental music clips; lip-sync and talking photos (experimental)
  • Prompt from an image, request preview, searchable and tagged job history with deletion, asset library
  • Template check before any GPU time is spent; licence tier on every feature; commercial-only mode
  • First-run Models & licences check naming each template's exact model files, licence tier and availability

In QA

  • JSON API with key protection

Planned

  • Installers for Windows, Linux and macOS
  • Commercially clean lip-sync and face identity; MuseTalk still needs dependency remediation and licence review
  • Signed Windows installer and winget package; ComfyUI Desktop and portable layouts
  • Verified behaviour on more graphics cards (8 to 24 GB)
  • Photo and text to 3D models (TRELLIS.2 research done; viewer already built)
  • Colour grading, trimming and joining clips, audio and voice, mesh clean-up, all as ComfyUI templates
  • Multi-subject animation from the prompt box
  • Optional adapters for Blender and, later, Unreal Engine
  • Not planned: cloud hosting or team seats. Plainbrush stays local and single-user by design.
The market

Where Plainbrush sits

GroupExamplesStrengthGap Plainbrush fills
Local enginesComfyUI, InvokeAI, SwarmUI, ForgeFree, private, the strongest open modelsExpert tools: graphs, checkpoints and settings to learn
Hosted enginesComfy Cloud, RunComfyNo setupPaid per use; photos leave the user's machine
Cloud creative suitesRunway, Firefly, Leonardo, Krea, OpenArt, Luma, Pika, Kling, Ideogram, CanvaPolished, broad featuresMonthly fees, uploads, opaque models and licences
Plainbrushlocal front end on ComfyUIPlain words, local, hardware-aware, licence-honest, checked resultsBreadth still growing; video and lip-sync are early

Competitor grouping from the project's own market research (28 Sep 2026); features and prices of other products change often.

Momentum

Eight weeks of building

17 SepStarted from a single fixed text-to-image workflow.
18 SepPluggable workflow templates; video as opt-in beta; real installer.
19 SepAutomatic object insertion into photos.
22 SepJob history and a real feature checklist.
23 SepConsistent characters across separate generations.
24 SepPhoto restore and upscale (SeedVR2).
25 SepDepth and sketch guidance.
29 SepTwo consistent people in one image; licence tiers and commercial-only mode.
30 SepRenamed Plainbrush; clean-install checks, CI on Windows, Linux and macOS.
1-5 OctDrag and drop, demo asset pack, photo editor that keeps the photo's shape, grounded object insertion.
13 OctVideo from a photo with whole-scene and face motion.
15-19 Oct3D direction researched; image quality gates on every result; classifier regression tests.
28 OctFunding and partnership scan: grants, accelerators and GPU-credit programmes shortlisted.
3 NovThis page went live; automated weekly test run (562 tests passing) and a weekly funding tracker.
6-7 NovGPU benchmark programme started: on the RTX 3060 the tensor-core unlock made SDXL images 49% faster and SD 1.5 images 36% faster (benchmarks).
12 NovSearchable and tagged job history shipped with safe row-only deletion.
13 NovFirst-run Models & licences audit shipped on desktop and mobile; test suite in progress with independent security and logic review. The first paid pack, store MVP, production pipeline and 11 Dec release gate were added to the roadmap.

On the development machine Plainbrush has run 545 jobs, 427 of them successful (78%), across 30 different workflows. None of the 118 failed jobs produced an image: they stopped with an error or a deliberate safety refusal (for example, not enough free graphics memory, or an experimental feature still switched off).

Where it goes

Roadmap

Now

shipping Natural-language front end for ComfyUI with 14 features, quality gates, licence tiers, a demo asset pack, Windows and Linux installers.

November

goal Fix the first Commercial Studio Image Pack scope and licence manifest by 20 Nov; start the store MVP on 23 Nov; continue published GPU-tier benchmarks.

11 December

release target Public beta, first verified paid pack and store MVP live together, with a real purchase/download/update test and a checksummed package on the Builds page.

Early 2027

goal Every feature offered only where the user's card can run it, backed by the benchmarks; a Blender adapter for 3D cleanup and turntables; Unreal Engine later as an optional pro adapter, built on the separate UnrealEngineAI project.

Funding

goal Applications prepared for the Mozilla Human-Scale AI Incubator (closes 20 Nov), Prototype Fund (30 Nov) and NVIDIA Inception, each decided and submitted by the founder.

Business

direction Free MIT core, with paid template and model packs, a bundle for studios, and a membership for new packs. Prices are proposals, not market data yet.

Packs and store

One release pipeline, one December gate

First pack

The Commercial Studio Image Pack is scoped by 20 Nov and verified by 4 Dec. It will include licence-checked image workflows, hardware profiles, real example outputs, documentation and acceptance tests. Unresolved face identity and lip-sync dependencies stay out.

Store MVP

Work starts 23 Nov after the SKU and licence manifest are fixed: product page, checkout, secure downloads, purchase records, versioned updates, customer licence terms, refunds and support.

Production pipeline

Research and official licence evidence โ†’ product selection โ†’ visual workflow spec โ†’ strict-TDD implementation โ†’ independent output, security, install and licence QA โ†’ GPU benchmarks โ†’ versioned, checksummed release.

Release gate

The goal closes only after a real purchase, clean install, download and update pass; the verified artifact and instructions appear on the Builds page; and Momentum records the release.

Before launch

Honest status

Product

Working daily on a home lab. Experimental features still have visible flaws (soft lip-sync, face drift in talking photos), and are labelled as such.

Name and company

The public site is plainbrush.io. A trademark filing, publisher name and company setup are still to do.

Licences

Several popular face and lip-sync models forbid commercial use. Those features stay off in commercial mode until clean substitutes are wired in.

How it is built

Most of the code is written by AI agents (Claude and a Hermes agent) directed by the founder, and the README says so openly. Some grant programmes exclude AI-generated work, so funding is aimed at investors and programmes that accept it.

Benchmarks

GPU benchmarks

The same ComfyUI jobs timed on each graphics card we test, stock and, where a tensor-core unlock exists, unlocked. First up: the RTX 3060 12 GB that runs Plainbrush today, where the unlock made SDXL images 49% faster (20.6 s to 13.9 s). See the benchmarks.

Funding

What support would do

Use of funds

  • Commercially clean model substitutes and the template packs that make the paid tier.
  • Test hardware across graphics-card tiers (8 to 24 GB) so every user gets the features their machine can run.
  • Company setup, trademark, legal review of model licences before the first sale.
  • The adapter layer for Blender and video editing.

Routes being prepared

  • Angel and pre-seed conversations, using this page and the live demo.
  • NVIDIA Inception and cloud startup credit programmes.
  • Early revenue from template packs before raising further.

Programme terms are re-verified on official pages before anything is submitted.