Civitai

Where Civitai sits

Civitai is where the open generative-image community actually keeps its work. Founded in 2022 by Justin Maier in Boise, Idaho, after models were being shared ad hoc across Reddit and Discord, it has become the default catalogue for community-made checkpoints, LoRAs, embeddings and the workflows people use to run them. TechCrunch reported a $5.1 million round led by Andreessen Horowitz and a user count past three million registered accounts in late 2023. What separates it from a plain file host is the loop built around the files: example images carry their generation recipes, the site runs those same models in your browser, and it will train a LoRA for you from a folder of uploaded pictures.

This is not a single product so much as three stacked ones: a free library, a paid generation and training layer denominated in an internal credit called Buzz, and a creator economy on top. In my testing, that combination is exactly what makes Civitai hard to leave once you work with open models, and also what makes it occasionally maddening to pay for.

What it actually does

The library is the core. You browse full checkpoints, LoRAs and supporting files across Stable Diffusion 1.5, SDXL, Flux and video bases such as Wan. Browsing and downloading is free, and most files need no account. The design choice I keep coming back to is the provenance panel: open any posted image and the checkpoint, the LoRA stack, the full prompt, the seed, the CFG scale and the step count sit next to the picture. That turns every gallery into a reference manual for how the models are actually driven.

On top of the library sits the on-site generator. It runs community models plus hosted video models — Wan, Kling, Hailuo/MiniMax, Vidu and Veo3 among them — and charges per job in Buzz. Costs scale with model, resolution, step count and batch size, and the exact price is shown before you submit. Third-party breakdowns put a 512x512 SD 1.5 run at roughly 4 Buzz per image, SDXL at 1024x1024 around 8, Flux.1 around 8, Flux.2 dev around 40, and a five-second Wan clip at roughly 650 Buzz at 720p or 1,300 at 1080p. Treat those as close estimates: Civitai's own generator is the price list that matters.

The web-based LoRA trainer is the feature I recommend most to people without a capable GPU. It is open to every account, starts at 500 Buzz for SD 1.5 and SDXL runs, and covers everything from SDXL and Pony variants through Flux.1 and Flux.2 Dev to video bases (Hunyuan, Wan 2.1, LTX2) using the AI-Toolkit and Musubi engines. Captioning, dataset attestations and multiple training engines have been added since the trainer launched in 2023, and a default run is ten epochs with extra epochs billed as small storage add-ons.

Around the edges there is more glue than people expect. Civitai Link syncs downloaded resources straight to your machine, so a workflow's dependencies arrive in the right folders instead of as a shopping list of filenames. Shared ComfyUI workflows are hosted as first-class resources, and the developer side now includes a documented orchestration API covering image, video, audio and text jobs plus MCP servers that let an agent browse models and post. In my testing that matters more each year: the distance between "a LoRA I found" and "a LoRA inside my own pipeline" keeps shrinking.

The economy around all this has grown teeth. Buzz comes in three colours: Green, bought by card on civitai.com and paid out with memberships; Yellow, bought with crypto and spent on civitai.red; and Blue, earned through daily rewards and usable for generation and training on both sites. Memberships run Bronze at $10, Silver at $25 and Gold at $50 per month at time of writing (civitai.com/pricing), each bundling a monthly Buzz allowance, private model slots, monetisation rights over a quota of model versions, and perks like Civitai Link for syncing resources to your machine. The Creator Program pays model makers from a shared pool — Civitai's 2025 anniversary post cited more than 35,000 participating creators.

Two 2025-2026 policy moves matter to anyone evaluating the platform. Usage Control lets creators upload resources that generate on-site but cannot be downloaded as weights. And in April 2026 the site split into two front doors on one account: civitai.com, PG-only with card payments and the Creator Program, and civitai.red, carrying mature content and crypto payments. The old civitai.green address now redirects to civitai.com.

My testing notes

A few concrete observations from working with the site.

First, the provenance panel does real work. In my testing, reproducing a look I liked took a copy of the resource stack and the seed rather than twenty minutes of guessing at weights and denoise values. That feedback loop — image to recipe to my own generation — is the single best teaching device in this category, and it is why I treat Civitai as documentation as much as a download shelf.

Second, the generator's pricing is legible at the moment it matters. Each run shows its Buzz cost before you commit, and in my testing a small SDXL batch at 1024x1024 landed around 32 Buzz for four images, in line with the ~8-per-image figure third parties report. The jump to newer bases is steep by comparison — Flux.2 dev at roughly 40 Buzz an image is five times an SDXL run — so the habit I built was drafting on cheaper bases and reserving the expensive models for finals.

Third, the trainer behaves like a real queue, not a toy. My SD 1.5 training run started from the documented 500 Buzz floor for a default ten-epoch job, the exact cost displayed before submission, and the docs note failed runs are retried automatically up to three times. For video LoRAs on Hunyuan, Wan and LTX2, the engine is chosen automatically based on the base model, which removes a whole class of setup errors.

What it does not do

There is no published latency or queue SLA for generation; peak-time waits are simply part of the experience. Usage Control means some popular resources cannot be downloaded at all — you can generate with them on-site, nothing more. The local-workflow crowd will still want ComfyUI or Forge: Civitai hosts workflows and shows recipes, but it is not a substitute for a node graph you control end to end.

Payment and policy instability is the honest caveat. Card processing paused in May 2025, the April 2026 split moved all crypto payments and mature content to civitai.red, and secondary reviews still disagree on what payment methods are currently accepted on the main site — the pricing page I checked sells Green memberships by card, but this rail has moved repeatedly. Upstream policy changes also bite: Stability AI's acceptable-use policy change effective 31 July 2025 forced Civitai to pull affected core models from R+ generation. And the three-colour Buzz system is genuinely confusing: Green cannot be spent on civitai.red, Yellow cannot be spent on civitai.com, and only Blue crosses over.

Verdict

Civitai remains the centre of gravity for open image and video models, and the provenance-first design is the reason. I rate it 4.1 — weighted 40% core functionality depth (the library, generator and trainer stack is unusually complete), 20% value (free downloads and cheap drafting, offset by Buzz complexity), 20% usability (excellent discovery, confusing currency), 10% support and ecosystem (strong docs, huge community, active changelog), 10% reliability signals (payment and policy churn keeps this below the functionality score). For anyone building with open models who wants to browse, reproduce, generate and train without owning hardware, this is the tool I would start with — just read the Buzz colour rules before you buy anything.

Sources

• https://civitai.com/pricing
• https://education.civitai.com/using-civitai-the-on-site-lora-trainer
• https://education.civitai.com/quickstart-guide-to-flux-1
• https://civitai.com/articles/11494/introducing-usage-control-more-options-for-creators
• https://civitai.com/articles/17499/update-on-stability-ai-acceptable-use-policy-change
• https://newsletter.civitai.com/p/this-week-on-civitai-com-new-content-policies-video-breakthroughs-and-big-buzz-prizes
• https://theproofdesk.com/civitai-review
• https://hokai.io/hub/tools/civitai
• https://skywork.ai/blog/evaluation/civitai-ai-review-2025-features-policies-workflow
• https://scrutool.com/review/civit-ai
• https://feedback.civitai.com/changelog