What Civitai publishes vs. what people actually use
I counted 2,224 LoRAs and 2,190 generated images on 31 October 2026. Characters are the most published type and the slowest to download. Styles are where the use is.
Snapshot date: 31 October 2026. This is a count of what is published, downloaded and used. It says nothing about which LoRAs are better. The platform moves fast, so treat every number as a photo of one day. The scripts that produced it are at the end, so you can take a new one.
Why I looked
I wanted to pick what to train next on a single 8 GB card. Training costs hours of GPU time, so I did not want to guess. The first thing I did was count what already exists.
All data comes from Civitai's public API, with no login, in Python's standard library only.
Measurement 1: what gets published and downloaded
What I collected. Five listings, which overlap, so 3,100 requests gave 2,224 unique LoRAs:
Most downloaded, this month: 1,000
Most downloaded, this week: 500
Most downloaded, all time: 500
Highest rated, this month: 500
Newest: 600
This is not a random sample of Civitai. Four of the five listings are rankings of what is already popular, and only "Newest" shows what is being published. Read everything below as "the top lists plus the newest uploads", not as "the whole platform".
Two traps I had to handle:
Download counts are cumulative. An old LoRA has thousands of downloads just from being old. Everything below is downloads per day since publication.
One republished model (548 days old) stretched the "Newest" window from under a day to over a year. I trimmed it. Without that, the publishing rate would have been wrong.
What came out. About 674 new LoRAs are published per day. Splitting by type, the Illustrious column is the one that caught my attention (160 new LoRAs in the sample):
Character: 131 new LoRAs (82%), median 147 downloads/day
Style: 17 new, median 803 downloads/day
Concept: 6 new, median 408 downloads/day
Clothing: 4 new, median 231 downloads/day
Character LoRAs are 82% of what is published and have the slowest median of any type. Style LoRAs are 17 of 160 and download about five times faster. The other ecosystems do not look the same: on MOSS Image characters are also slower than styles and concepts, but on Anima character and style LoRAs download at about the same rate (214 against 210 per day). The pattern is strongest on Illustrious, and I would not generalise it.
Measurement 2: what people actually generate with
Downloads are not use. Someone can download a LoRA and never run it. So I took a second sample: 2,190 freshly uploaded images, and counted which resources each one used. The 109 most-used resources were resolved to their name, type and base model.
Limits of this one, because they are big:
It is a 32-minute window of uploads. It depends on the hour and on whichever community had a busy half hour.
It only sees images that people upload to Civitai. Local generation that is never shared does not appear.
Without an API key, the prompt field comes back empty on 100% of images. You can count which resources were used, but you cannot read prompts.
What it showed:
Illustrious is 61.9% of the generated images and 26.8% of the LoRAs published. By that measure it is under-served.
Among LoRAs that people actually use, styles are 56.9% of uses. Illustrious styles alone are 48.1%.
The median image uses 4 resources. People stack LoRAs.
The 10 most-used LoRAs take 55% of all uses. The head is concentrated.
The three most-used Anima LoRAs were speed or distillation utilities, not content.
The sample did not agree with itself
I took two more samples of generated images sorted by reactions instead of recency. They gave an almost inverted order: MiniMax H3 at 34.6% of the weekly sample against 0.8% in the recency one.
That is not noise. Sorting by recency measures volume of work. Sorting by reactions measures attention. They are different demands. The only thing that held across all three samples was that Anima has more LoRAs published than any measure of use supports. I would not lean on any other ranking from a single sample.
A trap that invalidated my first answer
I tried to find "gaps": popular style LoRAs with no Illustrious version. The first pass found 4,361 of them. None was real.
Civitai's list endpoint truncates modelVersions. In my checks it returned 4 of 30 versions in one case, and in 5 of 6 suspicious cases it hid exactly the Illustrious one. Any question about which base models a LoRA covers has to re-request that model individually.
Once I did, the real picture was this. Of the top 150 style LoRAs, 100 lacked an Illustrious version, but 91 of those 100 were more than two years old, with a median of 39 downloads per day, against 155 per day for the 50 already covered. The gap was abandoned inventory, not unserved demand.
Also, "this model has no Illustrious version" is the wrong question. One popular Pony style had 154k downloads and no Illustrious version, and another author had already published the same artist's style for Illustrious. The unit that matters is the artist or style, not the individual model.
What I took from it
If you are choosing what to train, the type matters more than the base model. On Illustrious, characters are the crowded, slow corner; on Anima they were not.
Style LoRAs on Illustrious are where uploads and use line up best across both measurements.
Do not choose from one sample. Mine disagreed with itself.
Competing on quality in a proven niche is a legitimate answer. The data does not say "avoid it", it says "you will not find an empty space there".
What this does not tell you
Whether any given LoRA is good. Downloads and use are not quality.
The whole platform. It is the top lists plus the newest.
Anything about very restricted content. Without a login, the most restricted items are hidden from the API, so the share of mature content in the sample is a floor.
Anything next month. This changes fast.
Run it yourself
Everything is plain Python, no installs, and downloads its own data, so you start from today's numbers rather than mine. The scripts, with a README that explains each step and its limits, are in the ZIP of the guide below. Separately, and unrelated to the market data: once you have trained a LoRA, you need to know at what strength it works. I use a ComfyUI workflow for that, built only from core nodes (no custom nodes to install), that renders a grid of three prompts against five strengths with the same seed, so the only thing that changes along a row is the strength. It is attached to this article as lora_strength_grid.json: load it in ComfyUI, pick your checkpoint and LoRA, type the trigger word, queue.
This article is adapted from the first chapter of a longer guide I wrote about training a LoRA on a graphics card with 8 GB of VRAM: install, dataset, the configuration that actually finished, and what breaks. It is a paid guide, in English and Spanish, with the scripts included: https://aguireda.gumroad.com/l/lora-8gb-vram