Co-authored with Claude Sonnet 4
July 31st, 2024 marked the beginning of an unexpected adventure. After just three days of fumbling around with ComfyUI and AI text-to-image generation, I uploaded my first image to CivitAI. Looking back, I didn't even know how to use LoRA at that point! Little did I know that this tentative first step would launch me into a year-long journey that would see me create 1,000 images, 18 videos and 2 LoRAs.
First off, huge thanks to the CivitAI community for all the support โค๏ธ โ the reactions, comments, collections, and discussions really kept me going. As someone who tends to hop between interests pretty quickly, I probably would've given up after a couple months without you all.
The journey started with me messing around with Animagine XL V3.1 and other SDXL checkpoints. I tried getting fancy with ControlNet and regional prompter, but honestly, it felt like too much work, so I simplified things down to just txt2img plus upscaling. Then Flux Dev came along and totally changed the game โ I went through at least 208 different LoRAs with that model. Fun fact: it also turned my computer into a space heater during winter, which was actually pretty nice! ๐ Come spring though, I started noticing how much better the closed-source models were getting, so I made the switch. Plus, my fans finally got a break.
Now let me share some personal thoughts on the models I've used:
SDXL Checkpoints
Pros: Tons of style and character LoRAs to choose from, and they could output 1M pixels straight up (which was pretty impressive back then)
Cons: The model parameters are just too limited โ you can't really expect a 7GB model to do everything. This is also why I never bothered trying Illustrious XL despite all the hype on CivitAI. Plus, I really don't like its brush strokes, especially when it comes to rendering clouds.
Flux Dev
Pros: Great finger rendering (no more "bad hands/fingers" in negative prompts!), works well with natural language prompts, solid rendering of English characters, and plenty of LoRAs available
Cons: The infamous "flux butt chin" issue (though many LoRAs can fix this), banding problems (those horizontal and vertical dark streaks that upscaling can help with), and weird proportion issues between people and objects โ like when someone's sitting in a boat or relaxing under a pavilion, Flux tends to make people way too tall and oversized.
Those were the open-source models. Now let me talk about closed-source ones:
GPT Image 1
Pros: You can have multi-turn conversations to gradually refine your results, it can reference uploaded images, and it's incredibly knowledgeable about obscure concepts. For example, it knows how to render "ๆๅคๅคฉ็ฎ็" (ใใใธใใฆใใใใกใใใ, https://ja.wikipedia.org/wiki/%E6%9B%9C%E5%A4%89%E5%A4%A9%E7%9B%AE%E8%8C%B6%E7%A2%97) โ those rare tea bowls with oil-spot glazes that most models would have no clue about.
Cons: Way too trigger-happy with content policy violations (the main reason I gave up on it), and it just can't generate sharp, photorealistic images no matter how much you ask it to be more photorealistic.
Kolors (Kling's txt2img counterpart)
Pros: V2.0 was a huge leap for this model series โ it's the most "artistic" model I've ever used, absolutely excelling at ink wash paintings, watercolors, and thick coating styles. The recent V2.1 update focused on optimizing portrait photography and photorealistic rendering.
Cons: When using V2.0 for photorealistic images, prompts with "4K," "8K," or "ultra detailed" would often trigger a CG-rendered look instead (V2.1 largely fixed this). However, I feel V2.1 doesn't perform as well as V2.0 when it comes to anime and digital painting styles.
Reve Image (not included in CivitAI's tool list)
Pros: Pretty decent creativity, automatically optimizes your prompts (sometimes helpful), and gives you 20 free credits daily (20 images)
Cons: The portraits it generates are kind of weird, especially Asian faces, and hand rendering is problematic.
Seedream (another one not in the tool list โ used to be called "jimeng" before the UI update but got removed after)
Pros: The 3.0 version (integrated in Doubao) offers tons of style options, each basically acting like a LoRA, which gives closed-source models some much-needed style diversity. The recently updated 3.1 version (in jimeng) improved artistry and can easily generate great digital illustration pieces.
Cons: If you generate multiple images with the 3.0 version, you'll notice it has a strong tendency to produce the same face over and over โ making it pretty obvious when someone posts a Seedream 3.0 image that it's AI-generated. Version 3.1 improved this issue, but now when you want photorealistic images, it often gives you hand-drawn style instead and struggles to follow your prompts.
Imagen
Pros: I'm saving Imagen for last because I think Imagen 4 is currently the strongest txt2img model out there, especially for photorealistic images. Imagen 3 was already pretty solid at photorealistic generation, but you could only create 1:1 aspect ratio images in Gemini. With Imagen 4, it's made a spectacular evolution โ for photorealistic portraits, whether it's body details, lighting, perspective, or overall detail rendering quality, it's hands down the best performer in my opinion (you can really see this when comparing it to Flux 1.1 Pro Ultra).
Cons: In Gemini conversation sessions, you'll most likely only get one image per generation, even if you ask for multiple. Also, its viewpoint and composition variations are pretty limited โ after using it for a while, you'll find the images are technically correct with few flaws, but they can feel a bit boring.
That wraps up my thoughts on the different models I've used (I didn't mention Pony, Illustrious, or HiDream-I1 Dev since I haven't really used them much).
Personal Favorites (by Month)
Now let me pick my most satisfying work from each month's uploads:
August 2024
(Animagine XL V3.1,ย https://civitai.com/images/23665470)
Honorable Mention: https://civitai.com/images/27158449
September 2024
(Flux Dev, https://civitai.com/images/29507223)
Honorable Mention: https://civitai.com/images/27630451
October 2024
(Flux Dev, https://civitai.com/images/33697987)
Honorable Mention: https://civitai.com/images/35383744
November 2024
(Flux Dev, https://civitai.com/images/40744035)
Honorable Mention: https://civitai.com/images/38590901
December 2024
(Flux Dev, https://civitai.com/images/43755523)
Honorable Mention: https://civitai.com/images/43183944
January 2025
(majicFlus, https://civitai.com/images/51543973)
Honorable Mention: https://civitai.com/images/54313274
February 2025
(Flux Dev, https://civitai.com/images/59800077)
Honorable Mention: https://civitai.com/images/59367120
March 2025
(Flux Dev, https://civitai.com/images/61546974)
Honorable Mention: https://civitai.com/images/64784760
April 2025
(GPT Image 1, https://civitai.com/images/67406841)
Honorable Mention: https://civitai.com/images/67224728
May 2025
(GPT Image 1, https://civitai.com/images/74915550)
Honorable Mention: https://civitai.com/images/73429387
June 2025
(Imagen 4, https://civitai.com/images/82112179)
Honorable Mention: https://civitai.com/images/84574187
July 2025
(Seedream 3.1, https://civitai.com/images/91531467)
Honorable Mention: https://civitai.com/images/87864267
Wrapping Up
To wrap things up โ I'm a pretty deadline-driven person, and to hit my goal of posting the 1000th image on July 31st, 2025, I definitely ramped up my posting frequency these past few weeks. After this milestone, I'll be significantly slowing down my upload rate. Part of it is that I've picked up other hobbies that demand my time, but honestly, I also feel like I can't generate anything that really excites me anymore with current text-to-image tools โ especially in the portrait mode (like 3:4) that CivitAI tends to favor. I'm not planning to switch to landscape mode either. Current AI has gotten pretty damn good at single-person portraits, but multi-person scenes still feel like AI's weak spot to me (though I know XL users might disagree).
But this doesn't mean I'm leaving the platform for DeviantArt or somewhere else. Despite CivitAI's many issues, I just treat it as my online portfolio site, and it does that job well enough.
Anyway, I've rambled on long enough. Thanks again for all your support and love! ๐