Stable Diffusion 1.5 Cheat Sheet - Documentation & FAQ

Image Generation for Styles

All images were generated with either the Deliberate v2 or the DreamShaper 3.2 Checkpoint.
Sampler was DMP++ 2M Karras or DMP++ SDE Karras, depending on the better result.
Final resolution never smaller than 768px × 768px.

The prompts for the textbox were:

--prompt style of ArtistName --prompt style of ArtistName, woman --prompt style of ArtistName, Henry C_____ --negative_prompt superman --prompt style of ArtistName, city or village or landscape

While Deliberate has some NSFW Checkpoint mixed in and needs 'nsfw' in the negative prompt, DreamShaper likes manga a lot and frequently requires 'manga, anime' in the np.

Why these prompts?

The first line produces a broad, uninfluenced impression of what is most common for that artist (landscape, person, etc.).

With 'woman', you will get a wider range of portraits that are more in line with the artist's personal aesthetic.

For the third line: The actor Henry C. is currently well-known and has been well-trained into SD. You can also judge the dominance of a style if he is still rendered normally (weak artist style) or when his superhero costume appears.

The last line, 'city' (sometimes replaced by 'village'), will give you a good idea of how houses, people, and technical items (cars, bikes, etc.) appear.

Negative Prompt Styles

These are the negative prompts that were used to generate the images. They are reduced to contain the minimum keywords (human anatomy) to not influence the styles too much.

I'm including these for documentation purposes only; please make sure to do your own research.

Basic Negative Prompt mutation, deformed, deformed iris, duplicate, morbid, mutilated, disfigured, poorly drawn hand, poorly drawn face, bad proportions, gross proportions, extra limbs, cloned face, long neck, malformed limbs, missing arm, missing leg, extra arm, extra leg, fused fingers, too many fingers, extra fingers, mutated hands, blurry, bad anatomy, out of frame, contortionist, contorted limbs, exaggerated features, disproportionate, twisted posture, unnatural pose, disconnected, disproportionate, warped, misshapen, out of scale

'NSFW', 'Border' and 'Signature' were always selected.

NSFW nude, naked, nsfw, nipple, genitals, penis, vagina
Border border, frame, picture frame
Signature logo, text, signature, watermark, copyright

In order to limit the influence of manga/anime from checkpoint merges, it is occasionally necessary to include Southeast Asian countries in the negative prompt. This is not done for Asian artists or manga or anime styles.

No Anime / Manga anime, manga, asian [OPTIONAL IF NEEDED: chinese, japanese, korean]

How To Check if an Artist Is Known

Use the 1.5 base model of SD, which has all information uninfluenced by any merges, and look for consistency in the results. At the same time, check the search results for that artist's name to compare and get a feel for what they are known for.

Here are two examples, both made with the standard prompts:

Charles Dauphin - Two batches of four prompts, at 512px × 512px

The results for 'Charles Dauphin' are all over the place - photos, paintings, a catalog page - save to say he's unknown to SD.

Petros Afshar - Two batches of four prompts, at 512px × 512px

The results for 'Petros Afshar' are consistent. Similar colors, shapes, the art medium - SD knows this artist.

Sometimes things aren't as black and white, and more testing is required. Create a few more test images or switch to a more refined checkpoint/model and test there.

Forcing Results

Sometimes the generated images of a person are 'unflexible' or the style is not strong enough to produce the results you want.

Emily Balivet - The generated image of H.C. is not in the desired style

In this case, you can try and force a style by using the [from:to:when] tag.

[from:to:when] - [man:Henry Cavill:0.3]

This will begin with an image of a man in the desired style and will then transition to H.C. at 30% of the generation. (For example, if you set it to 100 steps, it would switch over at step 30.)

A good percentage for the switch seems to be between 0.25 and 0.35.

Additional Tips


Help! All my liked styles are gone!?

Your likes are saved in your browser and connected to the folder where the cheat sheet is saved. Simply move the folder back if you moved it.

How can I add my own styles? Why is there no option to add my own?

There is no simple way to add your own styles. Either manually edit the JSON data, or write your own script to edit the dataset.

I made a very simple PHP script for my own use, but I don't have the time to deal with any security or support issues if I made that publicly available.

Why no website?

I liked the concept of a tool that doesn't require the internet and runs on every computer (with a browser). No matter what happens elsewhere, all you have to do is download it, and it is yours. Renting web space that is quick and has a sufficient amount of traffic would also require financial investment. That's not something I'm interested in.

How much time did all of this take?

I'm using a Nvidia GeForce RTX 3060 (12GB) and generate a minimum of 20 images (easy to prompt style) and a maximum of around 120 images per style (if I can't get it right). All the free time! (I didn't do the math, sorry. :)