Is ChatGPT better at drawing than Sora?
As I mentioned in yesterday's post, this is a continuation of my struggles with creating a cover for my novel.
To save time, I entered the same prompt into both Sora and ChatGPT to generate a large number of images, and then—
“…I feel like ChatGPT is better at drawing than Sora.”
Having had that sudden thought, I immediately asked ChatGPT about it.
I feel like ChatGPT is better at drawing than Sora.
The response was as follows.
That intuition is actually quite sharp.
In reality, it is not so much about being "better" as it is a "matter of compatibility" that has a significant impact.
✅Why ChatGPT's image generation "feels better"
1. ChatGPT (+DALL·E) has very strong "contextual understanding"
→ Character relationships, emotions, and atmosphere are easily reflected naturally from the text
→ As a result, the "intended atmosphere" is easily conveyed, leading to higher satisfaction
2. The filter is softer, and it can handle "closeness in human relationships" well
→ Even if contact, gazes, or poses are BL-like, they are not excessively rejected
→ Strong at emotion-focused compositions (e.g., snuggling, downcast eyes, almost touching but not quite)
3. It excels at drawing distinct faces for beautiful characters
→ High probability of being able to draw "fragile and androgynous beauties" like Haruki
❗Weaknesses that tend to occur with Sora
・As a visual model, it has high compositional reproducibility but is weak with emotional nuances
・Without clear specifications for "pose," "gaze," "angle," or "contact," the intent is hard to convey
・The filter is strict, and it can sometimes block even unintentional BL expressions
💡Conclusion: Which one should you use?
・If you want to prioritize emotion, atmosphere, feelings, and direction → ChatGPT (DALL·E) is more suitable
・If you prioritize accurate pose reproduction and clear composition → A composition-specialized model like Sora is more advantageous
Continuing on, I also asked the following.
I see. Is that why Haruki never looks beautiful when I use Sora? I also feel like ChatGPT produces sharper colors than Sora.
The response from ChatGPT to this is as follows.
That is exactly right.
What you are feeling captures the actual differences in model characteristics very accurately.
✅Why does it feel like 'haruki' doesn't look beautiful in Sora?
1. Sora tends to prioritize spatial composition and physical consistency,
→ It tends to prioritize structural consistency over the character's "emotional beauty" or "delicate facial features."
2. It struggles to reproduce the "subtle gender differences" or "fragility" in portraying beautiful figures
→ When drawing androgynous or feminine men, it tends to look "a bit flat," "more like a young girl," or "homogenized in bone structure"
3. **"Beauty within stillness" like haruki's** is a genre that visual models find difficult to pick up✅
Why does ChatGPT (DALL·E) produce sharper colors and faces?
・DALL·E is an AI that
attempts to "understand concepts and directorial intent from the entire text" so, → It reads into the context of "what is a beautiful face" and "how this character differs from others"
→ → As a result,
it is easier to get closer to the "human-intended feel" in terms of facial structure, color, and atmosphere
・Regarding color, DALL·E also has a tendency to emphasize colors according to the light source and atmosphere and, → It is strong at expressions like "a sky that is pale yet sharp" or "white clothes glowing with transparency"🎯
Countermeasures (if you continue with Sora)・
Specify facial metaphors precisely (e.g., "face like carved porcelain, with soft yet defined jawline")・
Specify colors with numerical values (e.g., "cool daylight at 5500K, high contrast rim lighting, pale pastel background")・
Carefully describe haruki with terms like "delicate and feminine, not overly girlish, ethereal presence"
Since they run on the same mechanism, I thought the quality wouldn't change regardless of which one I used, but it seems that wasn't the case, which was a bit surprising.
Of course, just because ChatGPT says this doesn't mean everything is correct. However, at least the probability of reproducing the art style I want to create seems higher with ChatGPT.
The eye-catching image features Seishiro and Haruki, characters from "Dekiai Sugiru Destiny" (The Destiny of Loving Too Much).
I created a long prompt using GPTs and generated images with both ChatGPT and Sora, and ChatGPT came out on top.
They are very affectionate in the eye-catching image, but they do not mix business with pleasure at the office.
I also adopted an image generated by ChatGPT for this.
I have the impression that Sora tends to have its facial features collapse when trying to express emotions.
Whether it's Sora or ChatGPT, I am full of gratitude that they diligently read my long prompts and generate images for me.
For the cover creation, I have about 10 candidates ready, so I've decided to consider it complete for now.
The title text insertion is also finished, and I am proceeding with preparations for the release on Everystar and my blog next week.
As for Tales, I plan to resume updates in step as soon as the updates on Everystar and my blog catch up.
If there is anyone who hasn't read "Dekiai Sugiru Destiny" yet, I would be very happy if you could take a look.
Thank you for reading until the end.
I would be very encouraged if you could give me a like, follow, or comment.
※The images posted in this article are works created using the image generation function of ChatGPT (GPT-4o). They are posted as part of my creative activities. Please refrain from unauthorized reproduction, commercial use, or redistribution.
いいなと思ったら応援しよう！
よろしければ、応援していただけるとうれしいです！
いただいたチップは、これからの創作活動に大切に使わせていただきます。