How to Run Illustrious and SDXL Models Online Without a GPU

Key Takeaways

• You don't need a GPU. A hosted generator runs the checkpoint for you, and you only choose the model, write the prompt and set a few numbers.

• Illustrious prompts are tags, not sentences. The model learned Danbooru-style tags, so "1girl, silver hair, rain" beats a paragraph.

• Start with Euler a, 25 to 30 steps and a CFG near 5 to 6. Those are the numbers on the model cards we checked, and they're lower than vanilla SDXL's usual CFG.

• Licenses are per checkpoint. The base model's terms don't carry over to a merge, so read the license line on the exact checkpoint before selling anything.

Running Illustrious and SDXL without owning a graphics card

Illustrious XL is an anime and illustration model from OnomaAI Research, built on the SDXL architecture. It was trained on multi-level captions that mix Danbooru tags with natural language, according to its paper, and it handles a wider range of image sizes than vanilla SDXL. Most people run a fine-tune or merge of it, which is why Civitai lists so many Illustrious checkpoints.

Running one locally wants an NVIDIA card with real VRAM. If you have a laptop, a Mac or an older desktop, you can rent that card by the minute or use a site that has already loaded the checkpoints. This guide is written for illustrators and character artists who'd rather draw than debug CUDA, and it works the same way on any of the hosts in the list at the end.

Step 1: Pick a checkpoint and read its license

Browse Civitai or Hugging Face and filter by the Illustrious base model. The base release is Illustrious XL, with versions up to v2.0 on OnomaAI's Hugging Face page and v3.x on its own site, but most illustrators use a fine-tune for a more finished look. Open the specific version page, since each one lists its own recommended settings.

Check the license before you generate anything you plan to sell. WAI-illustrious-SDXL, a popular fine-tune, lets you sell the images but not resell the model. NoobAI-XL, another Illustrious-derived model, says in its published README that it prohibits commercialization. Hosted sites often show these as badges on the model page, and the point is that a base model's terms don't tell you about a merge's.

Stick with the standard (eps) checkpoints on a hosted site, since they're the ones every generator handles. Variants labeled v-pred, such as NoobAI V-Pred and Illustrious v3.0 and v3.5 VPred, need a generator that supports v-prediction, so skip them unless the site says it handles them.

Step 2: Open a hosted generator and set the canvas

Load the checkpoint in whichever host you chose. Illustrious v1.0 handles sizes from 512 by 512 up to 1536 by 1536, but the model cards for the fine-tunes we read recommend keeping the total area near 1024 by 1024. Safe starting sizes are 832 by 1216 for portraits, 1216 by 832 for landscapes, and 1024 by 1024 for square images.

Pick a size and stay with it for a few generations. If you change resolution and prompt together, you won't know which one moved the result.

Step 3: Write the prompt as tags

Illustrious responds to comma-separated tags in the Danbooru style, such as "1girl, silver hair, rain, city street, looking at viewer". The base model also accepts plain English, though tags are more predictable. Open with a few quality tags, then the subject, then the setting and style.

The NoobAI-XL model card prefixes prompts with "masterpiece, best quality, newest, absurdres, highres" and uses date tags (old, early, mid, recent, newest) to steer the look. WAI's card uses "masterpiece, best quality, amazing quality" and warns against piling on more quality tags or writing very long negative prompts. A short negative prompt such as "worst quality, bad quality, bad hands" is enough to start.

Step 4: Set sampler, steps and CFG

Euler a is the sampler both cards recommend. NoobAI suggests 25 to 30 steps and a CFG of 5 to 6, and WAI v17.0 gives 15 to 30 steps and a CFG of 5 to 7 with the VAE baked in, so you don't pick a separate one. Illustrious tolerates a lower CFG than vanilla SDXL or Pony, so if your results look burnt, drop it before changing anything else.

People argue about clip skip. The Illustrious cards we read don't state it and one community guide says 2, so follow whatever the checkpoint's own page says.

Step 5: Add a LoRA, but not three

A LoRA is a small add-on file that teaches the checkpoint a character, an outfit or a style. Most hosted generators let you stack several, but results degrade past two or three, and Mage's docs say so explicitly. Make sure the LoRA's base model matches the checkpoint, so an Illustrious LoRA goes on an Illustrious checkpoint.

Start with one LoRA at a weight of 0.6 to 0.8 and raise it only if the effect is too faint.

Step 6: Fix faces and upscale

SDXL models generate at about 1 megapixel, which is small for print or a wide banner. WAI's card suggests hires fix at 1.5 times with 20 hires steps, the R-ESRGAN 4x+ Anime6B upscaler and a denoise of 0.35 to 0.5. A lower denoise keeps the composition you chose, and a higher one repaints more detail.

Some hosts do this for you. Mage's SDXL Plus pipeline, for example, adds face and hand cleanup and a 2K hires pass, and it works with Illustrious checkpoints and LoRAs.

What else to know

• Free tiers have strings. Expect daily credit caps, queues at peak hours and, on some sites, watermarks on free output. Check before you build a workflow around one.

• Commercial rights follow the model. A host may allow commercial use on every plan, but an imported or community checkpoint still carries its own license.

• Hosted ComfyUI costs by the hour. Rented machines bill while they run, including idle time, so shut them down when you're done.

• Own-upload costs more. If you want a checkpoint that isn't in a site's catalog, Mage needs its $60 Pro Plus plan and ThinkDiffusion needs a rented machine.

Tools that run Illustrious online

Here are five places to run these models without a GPU, in alphabetical order. Prices are as of October 2026 and come from each site's pricing page where we could load it.

Civitai. It's the largest catalog of Illustrious checkpoints and lets you generate straight from a model page using Buzz credits. Some of its mature content moved to a separate site, civitai.red, in 2026, so read the current policy before you rely on it.

Mage. It's a browser-based generator with an SDXL catalog that includes Illustrious checkpoints, plus LoRAs, inpainting and negative prompts. Unlimited generation on base models starts at the $10 a month Basic plan, importing your own checkpoint needs Pro Plus at $60, and Mage's terms don't allow pornographic content.

SeaArt. It's a hosted catalog with Illustrious checkpoints and on-site LoRA training. Its free tier runs on daily credits, and reviewers report that free output carries a watermark and that the trial converts to a paid plan automatically, so watch the billing date.

Tensor.art. It's a browser generator with many Illustrious checkpoints and its own LoRA training. Free daily credits cover light use, and reviewers disagree about whether free output can be used commercially, so check the terms.

ThinkDiffusion. Hosted ComfyUI and Automatic1111 where you can load any checkpoint by pasting a Civitai URL. It bills by the hour, with Turbo 24 GB machines at $1.75 an hour, so it suits people who want full control and don't mind node graphs.