wk02 - MidJourney 2nd project

Post Reply
glegrady
Posts: 203
Joined: Wed Sep 22, 2010 12:26 pm

wk02 - MidJourney 2nd project

Post by glegrady » Fri Sep 16, 2022 8:03 am

wk02 - MidJourney 2nd project

Please post 4 to 6 images and their variations if you have any, with their text query. Following that, do an evaluation and commentary for each result. For this next step, explore any combination of techniques as documented at https://github.com/willwulfken/MidJourn ... /README.md and shared by Yixuan this past Thursday.

There is a wide variety of things to try out:

. The effect of accents over vowels in the prompts https://github.com/willwulfken/MidJourn ... parison.md

. The history of camera photography (angle of view, lens settings, etc.) https://github.com/willwulfken/MidJourn ... /Camera.md

The goal is to see to what degree one can control and discover
George Legrady
legrady@mat.ucsb.edu

wqiu
Posts: 14
Joined: Sun Oct 04, 2020 12:15 pm

Re: wk2 - MidJourney 2nd project

Post by wqiu » Tue Oct 11, 2022 11:08 am

Study 1: Sketch of multiple vantage points

Prompt: unusual perspective, hand drawn of Architecture
Vislab-MAT1_unusual_perspective_hand_drawn_of_Architecture_246c555c-32cd-4038-a7f1-48b6504564a4.png
This prompt generated a non-existent architecture. It looks interesting to me, an outsider to architecture.



Prompt: Vislab-MAT1_unusual_perspective_hand_drawn_of_Architecture_65915e5b-5b86-489d-b012-874581b4959a.png octane render
Vislab-MAT1_Vislab-MAT1_unusual_perspective_hand_drawn_of_Archi_8faea492-2fe3-4f8e-888a-84422920f8b9.png
Vislab-MAT1_Vislab-MAT1_unusual_perspective_hand_drawn_of_Archi_146d62c7-e90a-4e35-874a-3a21496b028c.png
This prompt accidentally included a long file name with lots of underscores and a string of hex code. It generated some interesting forms.



Prompt: unusual_perspective_hand_drawn_of_Architecture 65915e5b-5b86-489d-b012-874581b4959a
Vislab-MAT1_unusual_perspective_hand_drawn_of_Architecture_6591_7343bb7b-71b3-48f1-be76-a2aad6b92471.png
Vislab-MAT1_unusual_perspective_hand_drawn_of_Architecture_6591_385b078f-bbea-46ab-897f-8a1d011efa66.png
Removed "Vislab-MAT1_Vislab-MAT1" from last prompt and generated more reasonable results. Upscaled 3rd one.



Prompt: unusual_perspective_hand_drawn_of_Architecture_complex_structure
Vislab-MAT1_unusual_perspective_hand_drawn_of_Architecture_comp_c1f28f4f-42fb-4ecf-93af-c6a00e3e5bf7.png
Vislab-MAT1_unusual_perspective_hand_drawn_of_Architecture_comp_1b492341-d18d-4b18-88fd-f191b269063e.png
Added "complex structure" to the end, connecting with underscore to increase the correlation between the words.


Prompt: 65915e5b-5b86-489d-b012-874581b4959a
(text prompt from a filename affix, which looks like a random hex string)
Vislab-MAT1_65915e5b-5b86-489d-b012-874581b4959a_a98a7ea9-b7c7-4274-b2ff-a9ed20eba35a.png
Vislab-MAT1_65915e5b-5b86-489d-b012-874581b4959a_d64f67df-14cb-486b-b03b-5f47e5dbcab2.png
Testing the hex string by itself. The image looks like astrophotography. Probably the hex code text prompt looks like a space coordinate in the universe. Visually it looks pretty flat. The area of colors was organized diagonally, with tiny sparking stars distributed in the middle and bottom left. I prefer the small version to the upscaled version. I find the upscaled image looks less authentic.

Overall: 3.5/5


Study 2: fine art sculpture photos

Prompt: sculpture by tony cragg, photography by Edward Weston
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_8f61b204-ee32-4fa1-82c8-52d4a1ab5945.png
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_2198da24-5235-44f9-9d56-5c816ab7d9ed.png
Prompt: sculpture by tony cragg, photography by edward weston, top-down-view
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_858fbae7-39bc-4b44-ae7f-61b23fd02381.png
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_f0a97b54-875c-4d8a-800b-6149a3c61acc.png
Variations:
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_b444fe77-910f-418d-b344-d510e74b1fde.png
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_52d43093-4d56-4990-b46f-e991a78089bc.png
Vislab-MAT1_sculpture_by_tony_cragg_photography_by_edward_westo_a0a581ce-0f3e-4c0a-90c1-565782d1e797.png


Experiment with materials:
Prompt:black sculpture by tony craigg, fine details, marble texture
Vislab-MAT1_black_sculpture_by_tony_craigg_fine_details_marble__c4d892c0-1030-4f94-8fd9-d1aec1796ad2.png
Vislab-MAT1_black_sculpture_by_tony_craigg_fine_details_marble__e3d90eb6-8045-4c5d-8e2f-35b96568c843.png
Prompt: black stone sculpture by tony cragg, matt finish, photography by edward weston, Isometric projection
Vislab-MAT1_black_stone_sculpture_by_tony_cragg_matt_finish_pho_73aa220f-6c2b-444f-ab0d-79091347c45f.png
Prompt: black stone sculpture in vast desert, pure_black_annish_kapoor, form by tony cragg, matt finish, photography by Edward weston
Vislab-MAT1_black_stone_sculpture_in_vast_desert_pure_black_ann_c7cd70d3-e5d6-4256-89dd-12a78ddbf1b6.png

Overall: 4.5/5
I am surprised by the realistic texture applied to the sculptures. The forms of the sculptures looks interesting at the beginning, but it becomes boring after seeing all variants in similar style. I was hoping it can generate more style of sculptures. Maybe I should replace "Tony Cragg" with other artists. The other complaint is the finish of the sculptures. I tried to give the sculptures a matt finish but the keyword doesn't work. It might be due to the keyword "photography", because most product photo tagged with photography has this dramatic lighting effects which has a shiny highlight areas on the surfaces of the product.

tinghaozhou
Posts: 6
Joined: Mon Sep 26, 2022 10:24 am

Re: wk2 - MidJourney 2nd project

Post by tinghaozhou » Wed Oct 12, 2022 10:08 pm

This week I continue on exploring the question about how AI imagines and depicts speculative futures. Nnedi Okorafor's "Mother of Invention" continues to be the narrative foundation to which the Midjourney BOT keeps referencing. I started with this one simple description last week as a departure point for us to construct the imaginary world: "In the area between New Delta’s low skyscrapers, buildings and homes were carpeted with its world-famous stunning green grass, and the roads were fringed with it, but in this scene the grass was covered with smiley-faced bopping periwinkle flowers. It looked ridiculous, like one of those ancient animations from the early 1900s or a psychedelic drug–induced hallucination." I've come to realize that the BOT has a its own weight when trying to understanding and explicating a complicated, multi-word text prompts, which is possibly related to a statistically stronger association between particular terms in the prompt and the big data archive. That's why it seems that the BOT couldn't understand a sentence too well--we have to do the job by reducing (what we think) less important information in the prompt and help negotiate the relationship between data and noise in a prompt.

Another thing I noticed while experimenting with different configurations of a prompt was that the BOT seemed to understand the difference between a subject matter and a description of the style. For instance, when I used the prompt "Post-oil Nigeria, low skyscrapers, buildings and homes carpeted with greenest grass and smiley-faced bopping periwinkle flowers, ancient animations from the early 1900s," the term "Post-oil Nigeria" seemed to be manifested as the architectural form in the image and the "animations from the early 1900s" seemed to give the postcard-like hue to the picture.
Vislab-MAT1_Post-oil_Nigeria_low_skyscrapers_buildings_and_home_8f913eed-58ab-4c2e-90c3-6ac8b61d6cc5.png
However, I found the result not "futuristic" enough so I replaced the "post-oil Nigeria" to "afrofuturism": "afrofuturism, low skyscrapers, buildings carpeted with grass and smiley-faced flowers, animations from the early 1900s."
Vislab-MAT1_afrofuturism_low_skyscrapers_buildings_carpeted_wit_dc1d5f7e-20d3-457a-aaa0-54c67b0e373b.png
The image above was striking for me as the architectures are indeed more "futuristic" in the sense that it has a more postmodern form and shape and the flowers on the foreground are patterned in a way that helped create a seemingly "city of garden" in the future world. I consequently rendered a couple of more variations:
Vislab-MAT1_afrofuturism_low_skyscrapers_buildings_carpeted_wit_9aaea7d1-de72-4d0a-a40b-db902008a8a3.png
Vislab-MAT1_afrofuturism_low_skyscrapers_buildings_carpeted_wit_c968c164-79fb-4de0-8b86-c390a85b8dfc.png
The first image is fascinating as there is a human-like figure being created in the middle of the image (this imaginary world) even though the text prompt is rather "inhumane."
Vislab-MAT1_afrofuturism_low_skyscrapers_buildings_carpeted_wit_bc0052fa-9807-427a-adaf-11da63619342.png
Another variation gave out yet an even more detailed presentation of an architectural form that resembled the broadcasting tower we can see nowadays with a TV/screen/radar-filled overtop. Similar to this but with a reversed structure or even the Torre de Collserola in Barcelona:
Broadcasting tower.jpeg
Torre de Collserola.jpeg
However, as a variation of my original prompt, I believe the TV tower was only generated as a manifesting form of the keyword "skyscraper" as I didn't suggest tower or broadcasting in my prompt. These generated images are indeed more contemporary but not necessarily, "futuristic," from my own opinion; or to be more specific, they seemed to be futuristic in the last century. Looking at these pictures feels like reading a science fiction being written in the 1900s--I guess the prompt of "animations of the 1900s" really helped create this kind of atmosphere.

In the next step, I wanted my images to be more specific in terms of their styles--afrofuturism is a very broad term and a school of arts/thoughts that entail different artistic styles--because I feel like the generated images were more influenced by the keyword "animations from the 1900s" but not so much from "afrofuturism." Therefore, I decided to replace "afrofuturism" with two afrofuturistic artists to give a more signature form to the image. The first one is the foundational figure of afrofuturistic art, Sun Ra, who linked the African-American experience with ancient Egyptian mythology and outer space, creating early examples of afrofuturism in the mid-20th-century.
"Sun Ra, low skyscrapers, buildings carpeted with grass and smiley-faced flowers, animations from the early 1900s"
Vislab-MAT1_Sun_Ra_low_skyscrapers_buildings_carpeted_with_gras_48b4b77a-26d7-474a-b1b4-70f648973f34.png
The keyword "Sun Ra" instantly changed the style of the images, though still with similar subject matters (skyscrapers and grassland). Its explicit motif of the sun and the dominating yellow tone of the images successfully present Sun Ra's obvious heliolatry tendency but I'm not sure if they can actually represent Sun Ra's Egyptian tradition.
sun ra.jpeg
My assumption is that because Sun Ra is not a painter per se but a musician and an experimental theater performer, his artistic style is not easily represented in visual terms--the only references are his costume and prop designs which are pretty expressionistic. These images did give out the "surface-ness" of expressionistic designs but more likely because of the postcard style the "animations" prompt created.

The second figure is Wangechi Mutu. As one of the most important Afrofuturistic and feminist painters in contemporary art scene, Mutu's work explores the influence of Africa on other cultures, intertwining themes from her experience as an African woman and migrant. Although trained as a sculptor and multimedia artist, Mutu is best known for her large-scale collages, wildly colorful works including printed and painted papers as well as synthetic materials such as glitter or plastic pearls, and punctuated by armatures drawn in ink. A typical work of Mutu is in the style of a collage, often featuring a female body decorated with/mutated by botanical/floral elements:
Wangechi Mutu.jpeg
Hence I found it an even more perfect artistic style to depict an afrofuturistic story about pregnancy, mothering, distributed forms of care, and botanical futures.
"Wangechi Mutu, low skyscrapers, buildings carpeted with grass and smiley-faced flowers, animations from the early 1900s."
Vislab-MAT1_Wangechi_Mutu_low_skyscrapers_buildings_carpeted_wi_c4985c1c-e254-48fd-af0e-43e0f682cc1b.png
The images we got here are not necessarily in the form of collages but they definitely feature alienated forms of flowers and plants and large areas of light colors--pink, white, light green and yellow. Two specific variations are interesting though--presenting two possible directions of what the images could become: the first one is more pink and white, foregrounding the floral dimension of the prompt; the second one is more focusing on the architecture, surprisingly showing a human-like figure standing in the middle of a (seemingly) pollen storm.
Vislab-MAT1_Wangechi_Mutu_low_skyscrapers_buildings_carpeted_wi_7f27aa9c-3295-4cb9-bc0a-3d625ccff454.png
Vislab-MAT1_Wangechi_Mutu_low_skyscrapers_buildings_carpeted_wi_f18be040-90c3-4894-81c8-d45a87fdb211.png
Next week, I will build upon Mutu's style trying to re-tell/reimagine some sequences of "Mother of Invention." With different possible angles and pov prompts, I will try to explore the cinematic potentiality of these AI images.

jiarui_zhu
Posts: 7
Joined: Mon Sep 26, 2022 10:25 am

Re: wk2 - MidJourney 2nd project

Post by jiarui_zhu » Thu Oct 13, 2022 10:32 am

Last week I dedicated most of my effort exploring real, conventional images that MidJourney can generate. Natural scenes for example. One conclusion I reached is that MidJourney is able to generate all the elements in the text query and put them well together. However, even with the keyword “photorealistic”, most of the resulting images I got are all like painting style and lack the fine details.

This week, I want to have texts that are more imaginative and unusual. I want to generate images without an expected answer.

Text query:
abstract, hyper detailed, two opposite forces locked together and try to separate
Vislab-MAT0_abstract_Hyperdetailed_two_opposite_forces_locked_t_c07d2993-bd95-4be2-9cea-663bc084677b.png
I was very surprised by what I got. Force is a very abstract concept, and no one knows what it looks like in an image. The text query is also intentionally contradictory in nature. The resulting images use cold color and warm color to represent two forces, and they are all interlocked in different ways. The edges are a little blurry, which is a very good way of showing the “escape” concept in my opinion. The color choice is also opposite in nature. The angle of view is straight in front, like taking a photo of those images from a gallery.

I upscaled the third image to the max.
Vislab-MAT0_abstract_Hyperdetailed_two_opposite_forces_locked_t_9d38002b-7e7a-4113-a1b7-d8b3b861029f.png
I was very surprised by the amount of details and textures inside the image. The two forces look like they want to separate desperately so that the whole system is going to explode.

I also explored the effect of different styles.

Text query:
abstract, opalescent, two opposite forces locked together and try to separate
Vislab-MAT0_abstract_opalescent_two_opposite_forces_locked_toge_14d1bb00-3253-48a3-aa6e-8a17ce9d8424.png
Text query:
abstract, magical, two opposite forces locked together and try to separate
Vislab-MAT0_abstract_magical_two_opposite_forces_locked_togethe_97254414-4391-40e5-9f05-4741887f4911.png
I think the style texts have a very huge impact in setting the tone of the image. I am also surprised by how differing one word in the text query can change the resulting image so much.

I also up-scaled the third image to the max.
Vislab-MAT0_abstract_magical_two_opposite_forces_locked_togethe_0b2e56d5-efa7-465e-91ad-af8df55e87a0.png


Text query: everlasting, abstract, detail of light trying to escape the gravity of a black hole, gravitational wave, interlocked electrical and magnetic fields
Vislab-MAT0_everlasting_abstract_detail_of_light_trying_to_esca_40be3db0-da6e-4959-b2ee-0b3e8b9ac0db.png
The text query is also imaginative in nature. It consists of some physics concepts with adjectives that set the tone of the whole image.

I would rate this image with a score of ⅗. It's too noisy in my opinion and not abstract enough. The angle of view is again straight, much like taking a photo right in front of it. However, it does not show the oppression of being swallowed by a black hole.

I did some variations on the text with pretty much the same content.

Text query: everlasting, abstract, detail of light trying to escape the gravity of a black hole
Vislab-MAT0_everlasting_abstract_detail_of_light_trying_to_esca_98519107-0509-4040-b67a-ab0722cbc2d3.png
The first image looks better. I can sense the desperation of escape, so I upscaled it to the max.
Vislab-MAT0_everlasting_abstract_detail_of_light_trying_to_esca_70bfd16e-b344-403c-ad19-bf96193cc79e.png
Text query: everlasting, abstract, time and space distorted by huge gravitational field, universe
Vislab-MAT0_everlasting_abstract_time_and_space_distorted_by_hu_fb92d588-b714-4d63-a324-d4268cadb5d3.png
With the word “universe”, the image is more colorful, much like the stars and galaxies in the universe.

Text query: everlasting, hyperdetailed, severely distorted time and space in the universe, gravitational wave
Vislab-MAT0_everlasting_hyperdetailed_severely_distorted_time_a_22886f8c-7126-4765-b8fd-68028bcce047.png
This is probably one of my favourite images this week. I love how much details it presented. The distorted wave shape as well as the blurry starting from the center of the image give me a feeling of oppression. I am in owe.

Conclusion: I think MidJourney is better at generating imaginative unusual images than the normal conventional ones. With imaginative text query, MidJourney clearly has some understanding of what the text means and produces the images based on the understanding. Also style words have a huge impact on the overall tone of the image.

jkilgore
Posts: 7
Joined: Tue Mar 29, 2022 3:34 pm

Re: wk2 - MidJourney 2nd project

Post by jkilgore » Fri Oct 14, 2022 4:51 pm

Study 1: Halo, Drones, and Hands

Initialize

The first study consists of a single prompt, varied upon several dozen iterations. My goals was deep iteration to unveil local maxima the process hits upon.

It started with the following set of images and prompt:
iter0_halo.png
Prompt: playing halo, first person perspective, hands in front, death grips, realistic, Drone Photography, close up --chaos 60

'Playing halo' and 'realistic' significantly contributes to the futuristic, detailed, sci-fi look. Here is a screenshot from the game:
halo_example.jpg
From my previous study, I know that 'first person perspective, hands in front' contributes to hand like objects in the foreground.

'Drone photography' teased out warped landscape imagery. 'close up' was thrown in as a method for reeling in the wide angle tendencies of 'Drone photography'.

'Death grips' didn't seem to do anything significant.

First Stage

Hands, ships, structures reaching towards parabolic landscapes.
halo_parabolic_1.png
Second Stage

People emerge. Explicit showings of armor and helmets.
halo_ppl.png
Third Stage
The top left corner of the second stage leads to this blue dominate, closeup, synthetic look:
blue synthesis.png
Fourth Stage
The bottom right corner of the second stage leads to the following paths:
close ups and arches.png
Out of the top left corner, comes a split. One: textured closeups of vaguely hand figures. See it teased out in the top right. Two: more parabolic landscapes cut by blue shots of light. See it teased out in the bottom row.

Fourth.1 Stage
Beautiful textures on these ones. Some hands seem covered in crystals, others leather, others bark, possibly carbon fiber suits, similar to the ones depicted in halo.
half blue synthesis.png
more blue.png
Fourth.2 Stage
Some of the best framings I've gotten so far. The blue curves cutting through the scene provide a nice contrast. The geometries of the landscapes implied by the photos are hard to wrap your head around. Some images seem to depict the camera being under some large chunk of earth being lifted to reveal skies. Feels like a hurricane, but with soil.
para_scape.png
parabolic landscapes of blue.png

After Thoughts
Dedicating yourself to a single output and iterating upon it several times have consistently given me good results. It generates a certain level of depth not seen in doing 1 level deep prompts. It provides you a method for coercing the ml model into doing what you want through a guiding process that relies on aesthetic and curational skills rather than linguistic skills.

Study 2: Black Metal Concerts, Cameras, and Devolving Back into Hands

In this study, I attempted to be a bit more disciplined in my text queries. My goal was to see how I could affect the angle and stylize the shot based on some choice of camera. My subject was black metal and I stuck to one seed.

First Stage

Prompt: black metal concert --seed 2083
black_metal10.png
Second Stage

Prompt: black metal concert; super 8 --seed 2083
The super 8 adds a nice grain to the image. Great for black metal.
black_metal9.png
Prompt: black metal concert; iphone recording --seed 2083
black_metal8.png
Third Stage
Prompt: underground black metal concert; iphone recording --seed 2083
Midjourney takes 'underground' literally. In this case 'iphone recording' adds a nice warped look. I was attempting to pick up on images of people filming concerts with their smart phones.
black_metal7.png
Fourth Stage: Drones and Hands
Prompt: black metal concert; super 8; drone photography --seed 2083
The 'drone photography' keyword gave us a better chance of having 'above-angle' shots of the concert.
black_metal6.png
Prompt: black metal concert; super 8; fisheye lens --seed 2083
Midjourney took 'eye' literally...although it worked out in the end.
black_metal4.png
black_metal5.png
Fifth Sage: Hands (Best Yet)
Prompt: black metal guitarist in the desert; super 8; 2pm; closeup of hands --seed 2083
I fell prey to generating pictures of hands again... These are my best yet! The guitar got mixed up with the hand creating a strange cyborg hand guitar combo. The actual hands look the most real so far and the 'super 8' keyword adds a nice grain.
black_metal1.png
black_metal2.png
Attachments
black_metal3.png

lu_yang
Posts: 9
Joined: Mon Sep 26, 2022 10:23 am

Re: wk2 - MidJourney 2nd project

Post by lu_yang » Fri Oct 14, 2022 8:29 pm

The goal for this week is to explore how MidJourney can assist architecture design and how designers can control the design process. The tool needs to be able to produce consistent iterations from a single design and provide flexibility to control details in the image.

Initial test: brutalism, futurism, cyberpunk, interior, --seed 9878087
1-brutalism, futurism, cyberpunk, interior, --seed 9878087.png
1-1.png
MidJourney has produced a lot of interior views like the second one, which is my intention. But I really like the first one because it reminds me of Fritz Lang's "Metropolis", so I will proceed in this direction.
2-metropolis.png

Second iteration: I input the image from "Metropolis" as prompt and try to create a similar atmosphere and color tone. Interestingly, it produces stable results of interior views with different furniture setups and different glass facades.
https://s.mj.run/jhVv5UryRy8 ::reimagine, brutalism, futurism, cyberpunk, interior, --seed 9878087
3.png
4.png
However, a cyan tone is accumulated after a few more variations. This can be avoided by selecting the image with the right tone to iterate on.


Third iteration: I want to increase the interior concrete texture quality and lessen the reflective table surface, so I added "rough concrete" in the prompt.
https://s.mj.run/jhVv5UryRy8 ::reimagine, brutalism, futurism, cyberpunk, interior, rough concrete, --seed 9878087
6.png
two problems occurred during the process:
1. it is very hard to recreate the previous scene and the camera angle, even with the same seed number.
2. concrete texture was applied to the entire image, buildings far away not only change into concrete blocks, but also decrease in density and diversity.

Fourth iteration:
I used the image from the second iteration as prompt, hoping this iteration can inherit the original intent. Additionally, this time I will try a dusk/night view with urban lighting on the cityscape, while persisting concrete texture for interior space.
https://s.mj.run/LHYbIyz64QM brutalism, futurism, cyberpunk, interior, rough concrete, city night, --seed 9878087
7.png
8.png
The result is close to my expectation, however upscaling will create low-resolution and matte-style painting, and I believe this is because of the concrete texture being applied to the entire image.

In conclusion, the design process is painful and impossible to move forward, for a few limitations of the tool:
1. it is hard to use the same prompt to recreate the same scene.
2. the tool does not allow editing an existing image, which is the main reason why the design process cannot move forward.
3. word prompt is hard to relate to a certain area of the image. Designers usually want to adjust a targeted area, but word prompts cannot give clear instructions to the program.

merttoka
Posts: 21
Joined: Wed Jan 11, 2017 10:42 am

Re: wk2 - MidJourney 2nd project

Post by merttoka » Thu Oct 20, 2022 11:17 am

----------------------
You can see these explorations on a Miro board (Week 2) at this link: https://miro.com/app/board/uXjVPQgjXkQ= ... 6714658149
----------------------

To start this week's explorations, I want to share three images I liked. I used a modified version of George's prompt in class. I used one or two variations on some results. I liked the objects' translucency and crystal-looking nature and the images' overall composition.
ims.png

Study 1 - Camera Parameters

With this study, I wanted to see the effect of camera aperture and lens size on the generated images. I used one abstract and another tangible prompt for this exploration:

(for more information on this concept: Physics Project by Stephen Wolfram)
spatial hypergraphs --ar 16:9 --seed 2039
Vislab-MAT2_spatial_hypergraphs_ac5174d0-cde6-4244-89fd-58d330f77c25.png

a ceramic object resembling the forms in nature --ar 16:9 --seed 2039
image.png

Aperture
I ran the above prompts with various camera aperture values to see if the depth of focus would be reflected in the generated images.

(replace ### with f/ value of aperture)
spatial hypergraphs ### --ar 16:9 --seed 2039
sh_ap.png

a ceramic object resembling the forms in nature ### --ar 16:9 --seed 2039
ce_ap.png

Focal Length
I ran a similar set of prompts with various focal length values. This exploration produced better results because the imagery in the wide lens range contains more elements than in the telephoto range. However, this seems to be a one-off since the ceramic results don't reflect this property.

(replace ### with focal length)
spatial hypergraphs ### --ar 16:9 --seed 2039
sh_lens.png

a ceramic object resembling the forms in nature ### --ar 16:9 --seed 2039
ce_lens.png

I hypothesized that since the abstract imagery is not "shot" with a physical camera in the training dataset, it would have been challenging for Midjourney to correlate the physical properties of light passing the camera lens for these pictures. If this is true, ceramic images should have worked correctly. However, I cannot see either image's "correct" lens and aperture effect. I conclude that Midjourney doesn't understand the required technical camera specifications in the prompts.


Study 2 - Quality Parameter
The quality parameter determines the amount of time spend on generating an image. Its values are between 0.25 and 5, and the default value is 1. A value of 0.5 means it will stop the image generation at exactly half the time of the default generation time. Similarly, a value of 2 means it will spend double the time refining the generated image.

a ceramic object resembling the forms in nature --quality ### --ar 16:9 --seed 2039
Under-developed images (--quality 0.25-1):
ce_qu_01.png
Over-developed images (--quality 1-5):
ce_qu_15.png

I am more interested in under-developed images in this exploration than in over-developed images. Standard ceramics images generated by Midjourney look beautiful, yet it would be hard to 3D print if I wanted to design the form myself after getting inspired by Midjourney's outcome. However, I noticed that the under-developed versions of the same image have fewer crazy details, overhangs, and surface textures, so that they might be a better fit for physical fabrication.


Study 3 - Noise & Image Weight
In this study, I wanted to test (1) what noise means for Midjourney and (2) how the image weight parameter affects the results.

white noise --ar 16:9 --seed 2039
image (1).png
When I just use the text "white noise," I noticed Midjourney gives me ghosty humanoid figures.

IM --ar 16:9 --seed 2039
image (2).png
When I just use an image of white noise, Midjourney cannot perceive what is in the picture and generates another noise image.

IM white noise --ar 16:9 --seed 2039
image (3).png
When I use both the text "white noise" and the image of white noise, the initial ghostly humanoid figures become much more distorted, and the landscape becomes glitchy.


The last prompt is interesting because it lets us 'distort' the generated image. I wanted to see the range of control using the image weight parameter:

(replace ### with image weight)
IM white noise --iw ### --ar 16:9 --seed 2039
n_iw.png
As the image gains more weight, the prompt seems to lose its power and dissolve into complete noise after a value of 2. Values 3-4-5 produced the same result as the value of 2.

Post Reply