Home > Online Image Generators > NightCafe
Generated in DALL-E 3 at 7:4 aspect ratio in NightCafe.
To use NightCafe, just visit https://creator.nightcafe.studio/ and create an account. There is no phone app, but this website is useable from a phone. Just for joining you'll be given enough free credits to look around and try things, but some of the advanced functions require a pro account. I paid for the most basic plan which was $5.99 per month for the month I was testing it.
NightCafe is a community focused on image sharing, AI art contests, and group chat. To encourage people to participate in the community, you are given extra free points for things like liking other people's images, posting comments on a forum, or logging in on multiple consecutive days. NightCafe lets you create images using a choice of Stable Diffusion models or DALL-E. The main attraction of the service for me was the DALL-E 3 generation, discussed below, which includes advanced options such as a choice of aspect ratio which are not commonly found elsewhere.
In their latest update, now NightCafe has another leading model available: Ideogram. You can launch Ideogram or faster Ideogram Turbo generations from any paid NightCafe account. See the separate review of Ideogram for details.
NightCafe offers a selection of top SDXL models, such as Juggernaut XL, Real Clear XL, and Real Cartoon XL.
For LoRAs, NightCafe allows users to upload sets of images and train their own to use within their service. Anyone who wants to make art based on a particular person they have pictures of, or in a style shown in a set of images, will love making and using their own LoRA. NightCafe also offers a large selection of community-trained LoRAs to use in SDXL generations. (They call these "Pro models," meaning that you can only use them if you are a paid member.)
NightCafe offers an "advanced prompt editor" with a large menu of styles you can apply, which help you achieve different looks by adding words to your prompts.
There is also an option to upload your own images to regenerate (or evolve according to a prompt) with an image-to-image process.
The SDXL image generations will be approximately 1024 x 1024 in resolution (varying with aspect ratio.) Another image generation option is a workflow they call "Stable Core" that generates higher quality images approximately 1536 x 1536. Prompts are all screened by Stability AI's content filter, so this is not a place to create sexual or adult content, and there are occasional false positives where some innocuous prompt will be blocked by mistake.
After generating images, you have the option to upscale with algorithms such as REAL-ESRGAN x4. You can also "evolve" a generation by using it for image-to-image generation with a prompt.
Even though they don't offer every feature that you'd get if you had a Stable Diffusion interface running on your computer at home, this is a good representation of the core txt2img and img2img functions that you could start with. I hope in the future they add support for inpainting fixes and some kind of face detailing. Fooocus has proven that features like this can be implemented into an easy-to-use SDXL interface.
A real stand-out feature is that you can also enhance your images with "Creative Upscaler" which creates 4k resolution images, resampled according to your choice of prompts. I tried this just once and it looks like a powerful scaling option. Apparently, this is something new from Stability AI, and is not exactly like SUPIR or other modern creative upscaling workflows. Sadly, this new option is very expensive, costing 25 credits. (Normal image generation can be done for 1 credit, and my plan came with 100 credits per month.) I only tried it once and did not get to experiment with changing the "creativity level," which determines how much the image can change during rescaling.
I started with a fairly nice image of a beach house with a bike path beside it, ran it through the Creative Upscaler, and got a fairly nice image that was 4k resolution (over 4000 pixels in the longer dimension.) The results had some problem areas that I could clean -up with inpainting (in another program) if I wanted to perfect it, but overall it looked like a very powerful tool I'd like to use again.
After you generate an image in NightCafe, you are offered the option to turn your scene into a moving video with Stable Video Diffusion (SVD), for an additional 5 credits. SVD is very good at maintaining the look of your original image, so the first frame of your animation will look very similar to the frame that you started with. The animation SVD adds can be difficult to predict, however. If you put something in your scene with an obvious type of motion, such as water going over a waterfall or crashing surf at a beach, it will often give some nice motion. At times, SVD will create a camera move that makes your scene look more three dimensional. Overall, SVD is somewhat like a slot machine, where you sometimes get lucky, but the best results come from trying it many times until it happens to give you what you wanted.
I've already written a separate review of DALL-E 3, the leading AI image generator from OpenAI. Within NightCafe, DALL-E 3 images cost 4 credits each for default quality square images, going up to 8 or even 12 credits if you use the advanced options on the slowest settings. You get one image at a time, each time you type a prompt and click Create.
What's new and different in NightCafe is that there are advanced options available that I hadn't been able to use before. I tried using prompts that I had already used in DALL-E 3, so I knew I liked them, and focused on testing the advanced features.
The first advanced option is "runtime." It switches between "Medium" and "Long." The "Long" option costs four extra credits to use and seems to be what's called quality: "hd” in the DALLE-3 API. There was no option to lock the seed to the same number and see the exact same image at both quality levels, so I couldn't do A/B testing that isolated the difference well. Trying the same prompt with both settings gave me two different images, but I couldn't see any big difference that seemed worth the expense of leaving it on "Long."
The next advanced option is "Aspect Ratio." It offers a choice of 1:1, 4:7, or 7:4 aspect ratios. The default of 1:1 produces the square images with a resolution of 1024 x 1024 that we usually see from DALL-E 3.
4:7 aspect produces taller "portrait" images, with a resolution of 1024 x 1792. This is a welcomed option, as I frequently want vertically oriented images to share on social media, or to best capture pictures of people. It costs four extra credits per image, but gives you a taller, higher resolution image, and the quality of output can look terrific. Unfortunately, there were some recurring problems in generating the 4:7 images:
Some of the problem images are shown below, but when I generated enough of 4:7 images, I did get some terrific full-height images also. Review continues after break...
The 7:4 aspect ratio setting produces wide-screen or "landscape" images, with a resolution of 1792 x 1024. This option works great, and is well worth the extra 4 credits per image. You get higher resolution, bigger scenes that can look more cinematic than the default square images. A lot of the images in the photo album at the top were generated with this setting.
Another option, 'Style,' lets you switch between 'Vivid' and 'Natural.' When I first compared them, it seemed as if there had been some mistake, because 'Vivid' gave me natural-looking images, and 'Natural' gave me garishly oversaturated colors. Regardless of why they come out that way, the default setting of 'Vivid' provides well-tuned, predictable color response. The output I got with 'Natural' varied, but was sometimes oversaturated, and often overexposed parts of the image. Most scenes look better staying at the default.
Finally, there was an option called "Minimize Prompt Revision," which most likely sent an extra line of text to the large language model, saying something like "My prompt is long enough, please don't add any extra words to it." Results of using this varied, and without being able to use the same seed both ways, it was hard to tell whether the prompts were improved by this or not.
Of course, DALL-E 3 output is limited by OpenAI's content filter. On occasion a submitted prompt was blocked, but when this happened NightCafe did not charge me any credits for the creation, and I could just try again.
Copyright © 2024-2025 by Jeremy Birn
Welcome to the Internet! Websites use cookies to analyze traffic and optimize performance. By accepting cookies, data will be aggregated with all other user data.