Sign In

Hillside vs popular checkpoints, quality, sampler, photorealism

a_coffee_cup_with_tiny_sailboats_floating_on_the_coffee__surrounded_by_sugar_cube_icebergs___creative___masterpiece___unique_double_exposure_1.png
double_exposure_a_coffee_cup_with_tiny_sailboats_floating_on_the_coffee__surrounded_by_sugar_cube_icebergs___creative___masterpiece___unique_double_exposure_1.png
cinematic_still_a_desert_merging_into_the_sea___emotional__harmonious__vignette__highly_detailed__high_budget__bokeh__cinemascope__moody__epic__gorgeous__film_grain__grainy_-anime__cartoon__grap_1.png
cinematic_still_a_desert_merging_into_the_sea___emotional__harmonious__vignette__highly_detailed__high_budget__bokeh__cinemascope__moody__epic__gorgeous__film_grain__grainy_-anime__cartoon__grap_2.png
cinematic_still_long_shot_of_hkstyle_joan_of_arc__screaming_in_pain__arrow_stuck_in_shoulder__wounded_by_arrow__with_cropped_short_dark_hair__wearing_simple_medieval_armor__lying_on_grass__rainy_1.png
watercolor_painting_photo_of_old_book_with_text__joan_of_arc___old_book_cover_with_text__joan_of_arc___with_rough_watercolor_sketch_of_hkstyle_joan_of_arc_riding_on_horse__cropped_short_dark_hai_1.png

Here are several example images and thoughts for the images 1-6:

  1. I found an incredible image (by AIDigitalMediaAgency) and tried the prompt with Double Exposure LoRA. Result is good, but waves and table look like paintings on a very grainy canvas. This grainy effect happens too often, even when increasing steps and choosing different samplers.

  2. The sampler makes a huge difference. Same prompt as in 1, but "Euler A SubStep". Overall, "Euler A SYS" and "Euler A Trailing" worked best for me, and can lead to completely different results than "DPM++ 2M Karras", for example.

  3. I like photorealistic/cinematic pictures, but clouds look too apocalyptic. Tried different landscapes, lighthouse, castle, etc.

  4. For comparison: same prompt as in 3, but RealVisXL instead of Hillside.

  5. With some models, it is difficult to express sadness and pain. Hillside is subjectively better with uncommon facial expressions than several other popular SDXL models.

  6. Rendering of short text works surprisingly well. I needed three attempts to get this old-style book cover. With other models, I needed a dozen attempts to write "Joan of Arc".

And in general:

  • Hillside gives less consistent pictures when tuning some settings. Changing steps, sampler, LoRA weights, etc. just a little bit can give completely different. This is good to know if you'd expect something completely different from your prompt.

  • The model is a bit slower than others. However, if you generate individual images and want the best result, give Hillside a try. You might sacrifice a bit of quality ("grainy canvas"), but you're going to be surprised what it is capable of.

PS: For comparison regarding Joan of Arc facial expressions, you can find the full gallery created with RealVisXL here: