Civitai On-Site Generation: Settings, LoRAs and Buzz

You do not need a graphics card to use the models on Civitai. The site has an on-site generator that runs supported checkpoints and LoRAs on its own hardware and charges in Buzz, its internal credit. The controls are the familiar Stable Diffusion set: checkpoint, LoRAs, prompt, negative prompt, sampler, steps, CFG and seed. This guide walks through each one, shows a sensible starting point, and explains how to copy settings from an image you like.

Opening the generator

There are two ways in. From a model page, a generate button loads that checkpoint or LoRA straight into the generator. From an image, a remix or generate option copies its full generation data, including the prompt, negative prompt, settings and any LoRAs, so you start from a known-good configuration. Exact button labels change from time to time, but both routes exist in the current interface.

Not every model on the site is available for on-site generation. Models have to be in a compatible format and approved for the generator, so if a model page has no generate button, you will need to download it and run it locally. For that route, see the ComfyUI beginner’s guide or the local versus hosted comparison.

The controls, in the order to set them

1. Checkpoint. The base model. Everything else depends on its family (SD 1.5, SDXL, Flux and so on), so choose this first. The generator only offers LoRAs that match it.
2. Additional resources (LoRAs and embeddings). Add one or two LoRAs and set a weight for each. Start at the weight the model card recommends, otherwise around 0.7 to 0.8. Add the trigger word to the prompt.
3. Prompt. Subject first, then setting, then style and quality words. Trigger words near the front. Keep it under about 75 tokens for SD 1.5 and SDXL checkpoints; longer prompts are read in chunks and the later parts carry less weight. Flux checkpoints read natural sentences well and ignore most "masterpiece, 8k" padding.
4. Negative prompt. Things to avoid. A short list such as "blurry, low quality, deformed, text, watermark" is enough for most SDXL models. Flux models on the site may not use a negative prompt at all.
5. Aspect ratio or size. Pick from the offered presets. Use the native resolution of the checkpoint family (roughly 512 for SD 1.5, 1024 for SDXL and Flux).
6. Sampler and steps. The sampler is the denoising algorithm; steps is how many passes it makes. Use what the model card recommends. If it says nothing, a mainstream sampler at 25 to 30 steps is a good default. More steps past about 40 rarely improve anything.
7. CFG scale. How closely the model follows the prompt. SDXL-era checkpoints usually sit at 5 to 7. Flux models want a much lower guidance value, which the generator exposes on its own scale. Too high gives burnt colours and heavy contrast; too low gives soft, generic images.
8. Seed. Random by default. Fix it once you have a composition you like so you can change one setting and compare like with like.
9. Quantity. How many images per run. Generate four variations with a random seed when exploring, one with a fixed seed when refining.

Some checkpoints also expose a clip skip control or a VAE choice. Leave both at the card’s recommendation unless you know why you are changing them.

What Buzz pays for

Each generation costs Buzz, and the cost rises with image size, step count, quantity and whichever model you choose. The generator shows the cost before you confirm, so look at that number rather than guessing. Buzz can be bought, and the site also awards it for community activity such as posting images and receiving reactions; the exact rules change, so check the current help pages. The practical implication is the same as any credit system: iterate at small sizes and low quantities, and spend on full-size runs only when the prompt is proven.

Settings at a glance

For what each of these controls does under the hood, see Stable Diffusion settings explained.

Keeping and reusing your results

Every image you generate on the site keeps its generation data, and you can view it, download the image, or post it to the model page. The data lives with the image on Civitai, so if you download only the file you may lose the settings that produced it. Make a habit of copying the prompt, negative prompt, settings and LoRA list into your own notes when an image is worth keeping, and record the checkpoint version too, because models are updated and sometimes removed. A reliable test of your notes: can someone else reproduce the image from them alone?

Try it

Checkpoint: a photoreal SDXL checkpoint available in the generator
LoRA: none for the first run
Prompt: editorial product photograph of a matte black ceramic vase holding dried grasses, on a travertine plinth, soft window light from the left, neutral grey backdrop, 85mm lens
Negative prompt: blurry, low quality, deformed, text, watermark, cluttered
Size: 1024x1024. Sampler: card default. Steps: 30. CFG: 6. Seed: random, 4 images

Pick the best of the four, fix its seed, then add one style LoRA at 0.6 and run again. Comparing the two images with the same seed shows you precisely what the LoRA adds before you spend Buzz on anything larger.