Checkpoint Coverage on Civitai – A History
Since Civitai’s inception, one of our biggest challenges has been hardware limitations – the number of GPUs available to power services like LoRA training, image generation, and image tagging. While we’ve always wanted to support all Checkpoints for generation, our infrastructure simply wasn’t there yet.
For the past year, we’ve managed to support around 200 of the most popular Checkpoints for Image Generation, prioritizing widely used models to ensure the best experience for the most users. But the long-term goal has always been clear: full Checkpoint coverage.
Now, we’re finally at a place where we think we can make that happen! We’ve been:
- Rapidly expanding our own GPU resources and reducing reliance on third-party providers.
- Optimizing our Orchestrator and Worker code – the middleware that controls image generation.
- Building and deploying new systems designed to help scale up generation across tens of thousands of Checkpoints.
This rollout is a major step for Civitai, and while we expect some teething problems, we’re excited that we can finally make this a reality!
What to Expect: How We’re Rolling Out Full Checkpoint Coverage
To support tens of thousands of new Checkpoint models on the Generator while keeping our system efficient, we’ve introduced a popularity-based allocation system. This ensures that models with higher demand receive more compute resources, while lower-use models still remain accessible without overloading our infrastructure.
Every Checkpoint now has a “thermometer” badge that represents its current usage level, with seven levels:
Dormant > Quiet Buzz > Underground Hit > Getting Hype > Taking Off > Crushing It > Buzzing!
- Popular models get more GPU resources, meaning faster generations and a smoother experience.
- Less frequently used models are still available, but they will be allocated fewer GPUs to conserve resources.
Dynamic Pricing
To balance demand and availability, we’ve introduced dynamic pricing based on these levels:
- If a Checkpoint is “Dormant,” it will incur a +40% additional fee due to higher resource costs. Since these models see very little use, they must be loaded onto a GPU that could otherwise be allocated to higher-demand resources. This means the GPU is “tied up” for the duration of the generation.
- If a Checkpoint is “Buzzing!” its generation cost will be reduced by 20% thanks to optimized allocation. Since it’s in high demand, it will be preloaded across multiple GPUs, allowing it to handle multiple requests quickly and efficiently.
- The full breakdown is as follows;
| “Usage” Indicator | Markup/Discount |
|---|---|
| Dormant – Not currently being used for generation | 40% Generation Cost Markup |
| Quiet – Very low generation engagement | 30% Generation Cost Markup |
| Underground Hit – Some generation activity, but not frequent | 20% Generation Cost Markup |
| Getting Hype – Consistently generated with at a moderate level | 10% Generation Cost Markup |
| Taking Off – Frequently generated with, growing demand | No Markup or Discount |
| Crushing It – Strong generation demand, many users | 10% Generation Cost Discount |
| Buzzing! – Extremely high generation usage, peak demand | 20% Generation Cost Discount |
The popularity of a model is determined over time by a usage-based scoring system, meaning that as more people use a Checkpoint, it will move up the scale, leading to lower generation costs and better GPU availability.
This system allows us to fairly allocate resources, prioritizing the most-used models while keeping every Checkpoint accessible – even those that aren’t widely used.
Featured Models
Featured Models are a hand-picked selection of standout Checkpoints and LoRAs that appear prominently across Civitai, including:
- The homepage
- The Featured tab in the model picker
Being Featured gives a model increased exposure, making it easier for users to discover and use. But it’s not just about visibility – “Boosted” Checkpoint models also receive a fixed 20% discount on generation fees, making them more affordable for everyone to use.
Previously, Featured status was limited, static, and predominantly driven by Civitai, but now we’re introducing a new system that will allow the community to decide what gets featured – or “Boosted” – dynamically.
Auctions & Boosting
To make Featured status more accessible and community-driven, we’ve replaced the old Feature request system with a new Auction System that lets users bid on which models get featured.
Here’s how it works:
- Each day, 40 Featured slots are available for Checkpoints, along with separate slots for LoRAs (Flux, Illustrious, Pony, SD1.5, SDXL, etc.)
- Users can bid Buzz to push their favorite models into a Featured slot.
- Bidding runs for 24 hours, and the top bids at the end of the period determine the next day’s Featured Models.
- The top 3 models earn Gold, Silver, and Bronze status, receiving extra high-visibility placement across the site.
To participate, users simply navigate to the new “Auctions” link in the main Profile-Picture Menu, on the top right of the page.
- Select the model they want to bid on from the Model Picker.
- Enter a Buzz amount to bid, or use the quick-bid tools:
- “1st” – Auto-calculates the amount needed to push the model into first place.
- “Last” – Auto-calculates the minimum bid needed to secure a Featured slot (40th place).
- If the model has existing bids, users can add their Buzz to boost it higher.
Set and Forget: Recurring Bids
Want to keep a favorite model featured without rebidding every day? Users can enable Recurring Bids, which automatically re-bid a set amount each day until a chosen date.
- Recurring Bids can be paused or canceled anytime from the “My Bids” menu.
- This makes it easy for users to consistently support their favorite models without manually bidding every day.
Why Auctions?
- More Community Control – Featured Models are now chosen by users, not Civitai.
- Greater Accessibility – No longer limited to those who can afford a flat 10,000 Buzz fee; many slots will cost a lot less.
- More Dynamic & Fair – Models that users actively support get visibility, and all resources have a chance to be Featured.
This system keeps Featured slots fair, flexible, and open to everyone in a way that reflects community demand.
Frequently Asked Questions
Q: Does a model have to be Featured to be able to use it in the Generator?
A: No. All resources can be used on the Generator now, but Featured/Boosted Checkpoints provide a 20% Buzz discount for Generation, regardless of their popularity/Generation usage.
Q: Can PG-13+ content be Featured?
A: Yes, content of any rating can be Featured, but only PG content will appear on the Homepage. PG-13+ content will appear in the Featured tab of the Generation Resource Picker, subject to the level filtering applied by the viewing user.
Q: Boosted resources confer a 20% discount – does this stack with Usage based discounts?
A: No, 20% is the maximum discount a Checkpoint can have.
Q: Will there be Auctions for images? Articles?
A: Yes! We’re planning to expand the Auction system to give users full control over all featured content.
Q: Are all Flux finetunes available for Generation?
A: Not quite yet! The new Worker middleware we’ve rolled out does have support for these models, but it hasn’t yet been tested fully. We’re hoping to include full generation support for Flux finetunes in the coming days.
Q: What kind of Buzz can I use in the Auction system?
A: You can use Yellow Buzz to Bid on resources you’d like to see Boosted.
Q: How long are models Boosted?
A: Bids are placed for the following day. At midnight UTC, the top 40 models from the current day’s bidding are pushed into “Boosted” status for the following day, and the Auction system resets. And Bids which didn’t meet the minimum will be refunded.
