I'm running what amounts to an entry-level gaming rig. About a $1k outlay 18 months ago. Might be difficult to get anything below $2k currently. (This tech is going to be more widely accessible as the silicon-shortage hopefully eliviates in the next few years. But if you enjoy this kind of content, what's possible, etc., I'd recommend getting the software itself now, even if you can't use it. Get the various components [ComfyUI, Z-I-T model, NSFW LoRA, etc] so you can run it on whatever machine you have in the future. It'll be blazing-fast the way it is to run software from the 90s on a modern machine. Projections I've seen place that at 2-4 years.)
Asked two of the AIs I used (the well-known variety) to describe what I had them build.
====== LLM PRODUCT 1 ======
We've built a library of ready-to-use prompt templates for an AI image generator (Z-Image Turbo in ComfyUI).
Each template is a single text file for one theme (1970s catalogs, synthwave, classical rooms, gamer setups, and so on). Inside every file is a carefully written recipe that locks in the subject, hair, pose, camera angle, outfit, expression, setting, and lighting so the images come out consistent and high-quality instead of random.
The prompts are dense and structured on purpose so the AI has clear instructions on what to include and what to leave out. You pick a theme, paste the text, and generate. The whole thing is just organized prompt writing—nothing more.
--
**Here’s how someone with a [LLM product] membership can build the same kind of prompt library at home.**
### 1. Set up the basics
- Install ComfyUI locally and load Z-Image Turbo (or a similar model that responds well to dense, structured prompts).
- Create a simple folder on your computer for the templates (e.g., prompts/themes/).
### 2. Start a new theme with [LLM product]
Open a chat with [LLM product] and use a clear kickoff like:
> “New theme: [your idea]. Use the encoder playbook. Do not open the workflow JSONs.”
Then describe the look you want (era, clothing style, setting type, camera feel, etc.).
### 3. Iterate the same way the library was built
Work with [LLM product] to lock in:
- A consistent structure (numbered sections or the modular template style)
- Full, detailed packets for poses and settings (not one-line slots)
- Wildcard braces {option|option|option} for variation
- Clear rules for subject, hair, outfit, lighting, and camera
- Strong “what stays out” notes so results stay consistent
Keep refining until the prompt matches the density and reliability of the Christmas/Halloween examples. Ask [LLM product] to rewrite weak sections, add options, tighten the language, or test variations.
### 4. Save and use
- When a theme is solid, copy the finished prompt into a plain .txt file and save it in your folder.
- In ComfyUI, paste the text into the CLIP/text encoder node whenever you want that theme.
- Repeat for each new theme you want.
### 5. Keep a light playbook for yourself
Optionally ask [LLM product] to help you maintain a short personal notes file (syntax reminders, density rules, common fixes). That way you (or someone else) can stay consistent across many themes without starting from scratch every time.
That’s the whole process: structured conversations with [LLM product] to write and refine dense prompts, saved as simple text files, then used in your own ComfyUI setup. No special tools required beyond the membership and a local image-generation install.
====== LLM PRODUCT 2 ======
Here’s the short version of what we built.
The idea: a local viewer for the images ComfyUI makes, so you can browse, search, shuffle and favorite them, and it keeps itself tidy while you generate.
Three ways to look at your images
- Bucket is the main page. It has a grid, search, tags, favorites and videos, including a “local” video folder.
- Gridwall is a big, edge-to-edge wall of tiles. It can shuffle, show only favorites in random order, and search. Its menu is a compact pop-out by default, with an option to switch to a sideways row. A refresh button can stay pinned when the menu is collapsed.
- Catalog shows one image at a time, or two side by side, like a slideshow. It rolls new images on a timer, or follows the newest ones. ← and → step back and forward. There are optional old-TV looks (CRT and VHS, plus light versions that go easy on the GPU) and a No-Seam mode for the 2-up layout.
Find things by day: every page has a Day dropdown with â—€ Older / Newer â–¶ buttons. You can jump to “the day I did that theme” and leaf back through days. A day starts at 4:30 AM Eastern.
The server keeps things organized by itself
- New images get a date and time in their file name.
- Every 20 images, the finished ones are filed into webp/archive///. The old pile lives in archive/legacy/. This keeps the webp/ folder small enough to open for uploads.
- Your favorites and tags follow the files when they move, and a moved image no longer shows a broken tile.
- Every 20 conversions, the old raw PNGs are permanently deleted, leaving the newest one so the numbering stays continuous.
- A one-time script, refile.py, re-sorted the old weekly folders into the new daily layout.
Everywhere: all three pages read the whole archive, search the same way, and link to each other.
---- THE SERVER, built by LLM 2 ----
server.py is the small Python program you start. It does three jobs: it serves the pages and images to your browser, it converts new generations, and it keeps the folders tidy in the background.
Serving
- It runs a local web server on port 8000 and serves Bucket, Gridwall, Catalog, the images and the videos.
- It gives the pages their data through a few simple endpoints:
- the image list (it looks through webp/ and every folder under it)
- the video list (all subfolders, including local)
- your favorites and tags
- delete
- If a page asks for an image that has since been archived, the server finds it by file name and serves it from its new location instead of returning a 404.
Converting
- A background watcher notices new PNGs from ComfyUI in raw/ and converts them to webp.
- Each new webp gets a timestamp in its name, for example z-image-turbo_00585_20261002-104118.webp.
Housekeeping, every 20 images
- The finished webps move into webp/archive///, with the day starting at 4:30 AM Eastern. Anything loose at the top of the archive goes into legacy/.
- The old raw PNGs are permanently deleted, keeping the newest one so the numbering stays continuous.
- taxo.json, where your favorites and tags are stored, is rewritten to follow the moved files. It’s saved safely, so a crash can’t leave it half-written.
- Every few minutes it also checks that favorites and tags still point at files that exist.
|
