Seth's AI Tools: A Unity based front-end that uses ComfyUI and LLMs to do story and image/movie stuff
License: BSD style attribution, see LICENSE.md
Download the latest: V3.06 (Aug 30th, 2026) AI Tools Client (Windows, 150 MB) (codesigned by me)
- It's not a web app, it's a native locally run Windows .exe (Unity with C#)
- AI Chat mode is kind of like chatgpt/nanobanana, can just "ask" it to make images, posters, intelligently use anchor images, etc
- AI Chat can search the web (Brave Search API key needed, see Settings > Web) for reference photos and short video clips (bundled yt-dlp) and feed them to the image/video models, so "make a Seinfeld episode where they play retro games" can pull real likenesses; every search and download is shown in full in the chat. It can also read whole web pages (Wikipedia articles and the like) to research a topic and pull the pictures from a specific article as references, and a "Web" checkbox in the AI Chat header turns every online feature on or off
- Built-in image processing like cropping and resizing, mask painting
- Pan/zoom with thousands of images on the screen
- Built to utilize many ComfyUI servers and LLMs at once
- AI Guide feature harnesses the power of AI to create motivational posters, illustrated stories
- Adventure mode allows playing a realtime illustrated adventure (kind of like AI Dungeon) or generate quizes and choose your own adventures that can be exported to self contained playable HTML versions
- I've been using this as a personal AI testbed and tool/toy for OVER THREE years, weird experimental things are added and removed kind of randomly
- Privacy respected - does not phone home or collect any statistics, purely local usage. (it does check a single file on github.com to check for newer versions, but that's it)
- MiniMax H3 video generation: text-to-video, start-frame, reference-photo and reference-video modes with native audio and spoken dialog; turbo and quality tiers, now AI Chat's default video route
- H3 reference renders accept audio clips as voice/music style references alongside photos and video clips (up to 9 photos + 2 clips + 3 audio files per render)
- AI Chat audio generation (music, sound effects, speech) with Audio bubbles and set_video_audio soundtrack mixing
- AI Chat web research: search the web, fetch images/video clips (vision-checked), and read web pages for reference material
- Paste videos and files straight from the Windows clipboard; movie seek bar / playback fixes; video import previews with audio
- CLI can render H3 movies in all modes, including transparent-background VFX output (alpha webm / sprite sheets)
- Added Claude Fable 5 and GPT-5.6 model support
- ComfyUI servers can now be reordered in Settings to control which ones get used first
- Removed client-side LLM output token caps (model/server limits still apply)
- Fixed movies staying black after a mass reload on big canvases (ffprobe result caching, stalled-prepare watchdog, staggered reloads)
- Fixed AI Chat summarize bug and improved compact summaries
- Fixed text input not working right in the server configuration panel
- Added AI Chat video import/clipping with FFmpeg-backed transcodes, long-clip selection, preview proxies, audio controls, and still-frame import
- Added Bernini-R image editing/video-to-video presets and a separate RIFE video interpolation utility
- Improved video workflow reliability with MOV/HEVC processing proxies, source-duration Bernini frame caps, and unique ComfyUI output names for LTX renders
- Improved AI Chat attachment drops and fixed EXIF orientation on imported images
- Fixed Generate button and Unity 6000.5 build/editor issues
- Bundled FFmpeg/ffprobe helpers and licenses for video processing
- AI Chat now surfaces LLM/backend errors as visible bubbles instead of failing silently
- Added Opus 4.8, refreshed the Gemini model list
- AI Chat input box: draggable resize divider, scroll bar, and command-line style prompt history
- Stronger anchoring and multi-step "continue" handling so the model can carry out longer plans
- Comic creation now defaults to Ideogram 4, plus improved prompt caching
- Krea 2 Turbo image generation, AI Chat help system, background-removal and tiling skills
- Highlight words in AI Chat and have them spoken via ElevenLabs (add an API key)
- Force GPU reset option for reconnect, copy images to the Windows clipboard
- Tons of work done on the AI Chat feature as that's kind of all you need now
- New splash/startup screen
- Lot of tweaks and bug fixes, Ideogram 4 support added (AI Chat can directly write its json image description format for pretty good results)
- CrazyCam broken/disabled
- Things tend to get broken if I haven't used that particular feature in a while. It's probably the app, not you
Your ComfyUI server needs to be started using the --listen parm, so the API can be accessed.
After running it, a config.txt will be made. There is a config edit button inside the app, it's kind of self-documented, just go through it and update as necessary.
If your ComfyUI server is password protected, for example with the ComfyUI-Login custom node, append its direct API bearer token to the add_server line:
add_server|http://secured-box.lan:8188|token=$2b$12$qUfJfV942n...This is optional; leave |token=... off for normal open ComfyUI servers. For ComfyUI-Login, use the token it prints for direct API calls, not your web UI password. AITools sends it as an Authorization: Bearer header for ComfyUI HTTP and WebSocket requests.
To let AI Chat search the web for images and video clips, get a Brave Search API key at https://brave.com/search/api/ (the Search plan includes free monthly credit, roughly 1000 queries) and paste it into Settings > Web. YouTube and other page-hosted videos are downloaded with the bundled utils/yt-dlp/yt-dlp.exe (at up to 480p, then the requested seconds are cut out locally). YouTube needs a JavaScript runtime for its download challenge: install Deno (winget install DenoLand.Deno) or Node.js and the app picks it up automatically (Settings > Web shows what it found). If a site demands a sign-in, the same tab has a "cookies from browser" option (Firefox works most reliably). When a fetched clip will be used as a voice reference (the generated character talks), the app checks that the clip actually contains speech by transcribing its audio: point the tab's speech-to-text endpoint at any OpenAI-compatible Whisper server, or it uses your OpenAI key.
Tip: modern AI assistants (Claude, ChatGPT, etc.) can read this README and figure out how to download, install, and test the workflows AI Tools needs to run. I actually recommend doing it that way - just point your AI at this page and let it walk you through (or do) the setup, it saves a lot of time.
First, install ComfyUI and get it rendering stuff.
Then install my Workflow to API Converter Endpoint via ComfyUI Manager
This node allows us to only work with normal workflow json files, and never have to export "API" versions. The API conversions will happen under the hood automatically when needed, makes life much better. Click its page above for help.
It will work if you don't, but this allows modified or new workflows to work as it can convert them to "API" versions on the fly when needed.
Next, just for a test to make sure the workflows included with AITools are going to work, inside ComfyUI's web GUI, drag in aitools_client/ComfyUI/FullWorkflowVersions/text_to_img_flux.json or any others. The neat thing about ComfyUI is it will read this and convert it to its visual workflow format, ready to run. (you might want to change the prompt from <AITOOLS_PROMPT> to something else during testing here) - Click Queue. Does it work? Oh, if you see an Image Loader set to the file "<AITOOLS_INPUT_1>" you'll need to change that to a file on your ComfyUI server if you want to test.
You'll probably see a bunch of red nodes and errors - no problem! Make sure you have ComfyUI-Manager installed, you can use it to install any missing nodes. You'll probably have to track down various model files though, but at least when you try to render it will shows exactly the filenames that are missing. (look for red boxes around certain nodes)
Adjust it until it works (change paths or models or whatever you need, you could even start with a totally different workflow you found somewhere else), and make sure the prompt is set to <AITOOLS_PROMPT>. (<AITOOLS_NEGATIVE_PROMPT> can be used if your workflow has a place for that too) Then do Workflow->Export if you wanted to save your own. (note: we no longer use "API" versions at all, thank god, just normal workflows and dynamically convert them as needed)
So you don't have to create custom workflows for every checkpoint/filesize etc, you can use AITools' "@replace" to change any part of a workflow before it's sent to ComfyUI's API. You'll see it used in various presets.
Check discussions for some more info here
A small Python CLI in cli/ lets you generate images straight from a terminal, pointed at the same ComfyUI servers and Presets/ as the app. Great for scripting or for AI agents.
On Windows (the .bat auto-creates a venv and installs dependencies on first run):
cli\aitools_cli.bat "a cat playing a guitar" out.png -p "Prompt To Image (Z-Image)"
cli\aitools_cli.bat "make the sky red" out.png -p "Image To Image Klein Edit 1 Input" -i input.pngOn Linux/macOS:
python cli/aitools_cli.py "a cat playing a guitar" out.png -p "Prompt To Image (Z-Image)"-p picks a preset (any text-to-image or single-step image-to-image preset works; pass an input image with -i). Add -v for verbose output. It reads cli/config.txt for server settings - see cli/README.md for details.
AI Chat is a feature is kind of "but we have chatgpt image stuff at home" that allows an LLM to basically do everything - try things like "Make a funny motivational poster about the commodore 64", or attach three images of people and say "dress these three people like the 3 stooges". It will automatically use zimage, Klein based image to image, LTX, etc as needed to perform whatever actions you want. You can also drop .mov, .mp4, or .avi videos into AI Chat and use short imported clips for video-to-video edits.
It can intelligently create and use 'anchor images' as needed to provide consistency across multiple images, for example if you ask it to "make a comic book about my dog and cat" it can create an anchor image of the dog and cat and reuse that across panels to keep them looking the same.
It has an automatic "skills" system (it loads them as needed by itself) so it knows how to do things like create illustrated storybooks, comic books, etc.
An LLM as smart as Qwen 3.6 27B or better recommended, in addition to at least one ComfyUI server setup.
You only need to download the zip and run the .exe to use this, However, the source might be useful to generate a build for other platforms, fork or steal pieces to use for yourself. Go ahead!
- Requires Unity 6.4+
- Open the scene "Main" and click play to run
- Assets/GUI/GOTHIC.TFF and Assets/GUI/times.ttf are not included and might break the build because I was having trouble and switched some settings around that might require them now (dynamic vs static TMPro font settings...)
This project was developed with assistance from AI tools. I mean, you can still blame me (Seth) for bugs, but I just wanted to mention it.
- Audio: "Chee Zee Jungle"
Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 3.0
http://creativecommons.org/licenses/by/3.0/
- NotoSansCJKjp-VF font licensed under the Open Font License (OFL)
- Audio: JOHN_MICHEL_CELLO-BACH_AVE_MARIA.mp3 performed by John Michel. Licensed under Creative Commons: By Attribution 3.0
http://creativecommons.org/licenses/by/3.0/
- Written by Seth A. Robinson (seth@rtsoft.com) twitter: @rtsoft - Codedojo, Seth's blog
- Special thanks to the ComfyUI and llama.cpp projects