Expanding the Potential of AI: Challenges and Discoveries in Image Generation
The Challenge of the ‘Doing Something’ Theme
Merry Eve! It’s kojirom here, alone as always. This time, I decided to try out a theme called “doing something” for the first time.
Every day, as I ponder what to choose for the prompts, I sought advice from AI, and Claude suggested that a scene of “someone doing something” might be a good idea.
I thought this was a fantastic idea and decided to give it a try.
However, the results didn’t turn out as expected.
It’s difficult when you’re simply told to depict “the appearance of doing something” due to the lack of specificity.
As a result, many incomprehensible images were produced, and most of them ended up being discarded.
Therefore, I settled on the following theme for this time:
“A single girl” → “In the act of doing something” → “In a swimsuit”
Since there were many rejected images, their arrangement is random. I hope you can understand.
A-Zovya RPG Artist Tools – V4 | Stable Diffusion Checkpoint | Civitai
[Author = Zovya]
Breathe – v1.0
Breathe – v1.0 | Stable Diffusion Checkpoint | Civitai
[Author = TridIsntAName]
Bunnyspoon – Bunny4Spoon
Bunnyspoon – Bunny4Spoon | Stable Diffusion Checkpoint | Civitai
[Author = LyloGummy]
HoneySlime’s -Merge- (AnimeTestModel) – v1.0
HoneySlime’s -Merge- (AnimeTestModel) – v1.0 | Stable Diffusion Checkpoint | Civitai
[Author = HoneySlime5546]
Mistoon_Ruby – v3.0
Mistoon_Ruby – v3.0 | Stable Diffusion Checkpoint | Civitai
[Author = Inzaniak]
Radiant mix – v1.0
Radiant mix – v1.0 | Stable Diffusion Checkpoint | Civitai
[Author = Sas_ass]
SoteMix – v2.1
SoteMix – v2.1 | Stable Diffusion Checkpoint | Civitai
[Author = Disty0]
The Ani-Loop – 1.5
The Ani-Loop – 1.5 | Stable Diffusion Checkpoint | Civitai
[Author = darkseal]
Yet Another Anime Mix SXXVFDE – v2.0
Yet Another Anime Mix SXXVFDE – v2.0 | Stable Diffusion Checkpoint | Civitai
[Author = SXXVFDE]
A-Zovya RPG Artist Tools
It is enhanced for artists in the RPG industry and is suitable for creating illustrations of creatures, RPG motifs, etc.
V4 adds more illustration-style styles compared to previous versions, improving ease of use.
V4 incorporates training specialized for RPG elements as well as new illustration expression styles.
Various styles can be specified in prompts.
Default settings for each model can be checked in a table.
There are also multiple V4 variations such as V4 inpaint, V4+VAE, V3+VAE. Each has different applications and characteristics.
Stable Diffusion 2.1 (768) version is also introduced. It has high performance and high resolution but is not suitable for human figure expression.
We also accept support and requests from model developers.
(Site Summary)
“This is an advertisement.”
FANZA Illustration Collection (R18)
Exclusively available on FANZA!
My CG illustration collection for adults, created by me, kojirom, is now on sale to rave reviews.
Please check it out!
Breathe
Breathe is an AI model for generating images, combining 2.5D and anime aesthetics. It excels in creating realistic landscapes and attractive characters, offering tools for translating imagination and concepts into visual forms. It recommends specific negative prompts and settings, like DPM++ 2M Karras, 35 steps, CFG 8-12, and using highres.fix and img2img SD Upscale for model images. The author favors settings like 0.55 denoise and 20 hires steps, suggesting using Anime6B upscaler for SD Upscale and avoiding Restore Faces. These settings are personal preferences and may vary based on different artworks and environments.
(Site Summary)
Bunnyspoon
Bunnyspoon is an AI model designed for cuteness, aiming to be a favorite in HG and Civitai communities. The creator, an AI enthusiast active since Stable Diffusion’s launch, offers the model for free and appreciates any support through Ko-Fi or Patreon. The recommended settings include a step count over 20, DPM++ 2M Karras sampler, and specific scales and sizes. The model is free for personal use, but redistribution and commercial use have specific conditions. Users can find samples and a cookbook on Huggingface, with a disclaimer for using the model in mixes.
(Site Summary)
〈AD〉
Exclusive Sale at FANZA!
My CG Illustration Collection (for adults) by kojirom is now available and highly acclaimed.
Please check it out!
HoneySlime’s -Merge- (AnimeTestModel)
The model is in development and available for testing. It’s not fully functional yet and has several issues, but users are encouraged to try it and use it for their merging projects. Updates on the final version will be provided in the description section once it’s complete.
(Site Summary)
Recommended Article
Mistoon_Ruby
Subscribers to the Patreon can access all LoRAs, including exclusive ones, in a single package. ‘Mistoon_Ruby’ is a variant of the ‘Mistoon_Anime’ SD model, blending Western cartoon and modern anime styles. The model, which can create NSFW content, has been used for examples on HiRes-Fix, and a guide is available for more details. Preview images were generated with Ranbooru, scraping tags from various image sites. The creator also mentions their DeviantArt and Redbubble Shop.
(Site Summary)
〈This is an advertisement〉
Kojirom’s Adult CG Illustration Collection Now On Sale!
An adult CG illustration collection by kojirom is now available on FANZA.
Starting from just $1.0
It’s highly praised for its delicate touch and unique worldview.
Don’t miss this opportunity to check it out.”
Radiant mix
The creator combined models, was pleased with the simple yet effective results, and shared them. The mix is particularly effective for creating cartoon-style images, especially in rendering eye expressions. The creator looks forward to seeing others’ work using this mix.
(Site Summary)
SoteMix
This high-resolution anime image generation model (1024×1536) specializes in manga-style, particularly in eye expressions, and is not recommended for use at 512x resolution. It’s available on Hugging Face and supports OpenVINO and ONNX. The model offers multiple resolution options, but might be unstable at certain high resolutions. Key features include different samplers, steps, and upscalers, and it was trained with several resolution variations and model merges. Notably, it integrates technologies like SoteMix V1, Lunar Radiance Light, and Blessed2 VAE, and was trained on an Intel ARC A770 16GB GPU.
(Site Summary)
The Ani-Loop
※ I really liked this model, so I will feature it tomorrow. Stay tuned!
Yet Another Anime Mix SXXVFDE
The model combines various styles for flexibility, tending to draw 3D-like noses with ‘Detailed’ or ‘Realistic’ prompts, but typically renders classic anime noses (changed in version 2.0). It supports both SFW and NSFW content and requires a VAE. Users are recommended to use their preferred sampler.
(Site Summary)
All Illustrations Are Commercially Usable
Some illustrations may have incomplete details such as hands and fingers. We kindly ask for your understanding.
When using these images, please feel free to crop or modify them as needed.
We encourage you to fully express your creativity and use these illustrations with a free and imaginative approach.
The illustrations featured on this site are available for commercial use.
Stable Diffusion Web UI Settings and Parameters
Positive prompt: doing something,
Negative prompt: EasyNegativeV2,
Steps: 80, Sampler: Euler a, CFG scale: 6.5, Seed: 857405761, Size: 512×512, Model hash: 61d0cd0528, Model: radiantMix_v10, VAE hash: 0b204ad0ca, VAE: klF8Anime2_klF8Anime2VAE.pt, Denoising strength: 0.32, Clip skip: 2, Hires upscale: 1.5, Hires steps: 40, Hires upscaler: 4x-UltraSharp, TI hashes: “EasyNegativeV2: 339cc9210f70”, Version: v1.7.0
Postscript: Exploring AI Services and Their Surprises
As I mentioned in the beginning, I failed after following Claude’s advice. However, I have been using chatGPT-4 for a long time and I was satisfied with it.
However, I recently tried ChatHUB, a service that allows you to ask questions to four AIs (Google Bard, Microsoft Bing, Claude, Perplexity, etc.) at once, and I was surprised by the results.
In particular, I noticed that Claude’s performance is on par with the paid chatGPT-4, and Bard and Bing are also quite powerful.
After repeatedly asking the same questions to these five AIs, Claude seemed to be the best. However, the difference is slight, and the adoption rate of chatGPT-4, Bard, and Bing is also almost the same.
It is also interesting that the five AIs provide different codes in response to code requests.
Surprisingly (?), Bing provided the functioning code.
ChatHUB is a recommended service.
It is paid, but it is a one-time purchase, and at the current exchange rate, it is about 5,250 yen.
In particular, I would like you to try Claude even if you don’t want to spend money. It should be free to use individually.
If you only do chat, it may be possible to choose Claude. However, the existence of DALL-E 3, which can draw pictures, is large, so I will not switch from chatGPT-4 to Claude from next month.
In terms of overall strength, chatGPT-4 is still superior, but Bard seems to be the smartest in Japanese.
However, Bard often refuses to answer, which is stressful.
In that respect, chatGPT-4 is expected to have high user satisfaction because it will properly respond to any question.
This turned out to be a long post.
I will accept requests for models or requests from modelers to “introduce my model.” Please feel free to use this blog.
Then, I hope you all have a wonderful Christmas!
