Back to Discussions
Tips & Tricks

Hypothesis verification for generating high-quality anime-style images with Seedream

ねねむ
Translated to English·

I want to create great illustrations with Seedream! But even when I use credits to generate them, it doesn't always turn out well...

Have you ever had that experience? I have 😂

Plus, on this site, there are really few descriptions of Seedream’s illustration characteristics, and the only ones available are Maren’s Seedream personification works... (See here.)

So, I’m writing this article as a reference for effectively using Seedream to generate beautiful anime-style illustrations 🖋️

【What Are Seedream’s Characteristics?】

According to Maren’s observations, Seedream excels at "dresses, flowers, and delicate illustrations" and has "a soft, delicate touch as its feature"!

With that in mind, I analyzed Seedream’s anime-style illustrations on this site myself and observed the following characteristics:

① High resolution allows for fine-grained detail in intricate areas 🔍️
② Delicate color usage, good at creating a hand-drawn feel like watercolor paintings 🖌️
③ Excels at light depiction (gloss, reflections, transparency, etc.) ✨️

【Reverse-Engineering Prompt Construction】

Now, based on the above, I thought the following prompts would be effective for generating higher-quality images with Seedream:

① Prompts related to detailing 🔍️
② Prompts that define the style, such as "watercolor painting style" or "oil painting style" 🖌️
③ Prompts related to light depiction ✨️

Also, to verify whether quality-enhancing prompts like "highest quality" are effective for Seedream, I will conduct this comparison and test the hypothesis!
(Though Seedream doesn’t allow seed value fixing, making a precise controlled experiment impossible 😓)

【Experiment Details】

① Generate with the prompt: "anime style, watercolor painting paint application, girl, blue eyes with light particles, black hair glowing aqua, blue and black butterfly-shaped hair ornament, blue ribbon tie, blue and black gothic lolita-style outfit with lace and frills on the collar, hem, and sleeves, grassland, white and pink cosmos flowers in the foreground, shining autumn foliage trees in the background, emerald and aqua gradient sky, sparkling, lens flare, rim light."

② Next, generate with "highest quality, ultra-delicate, masterpiece" added at the beginning as a quality-enhancing prompt.

※The original artwork could not be uploaded due to file size limitations; compressed versions are shown here

↓Without quality-enhancing prompt
image

↓With quality-enhancing prompt
image

The background gradient in the second image is softer, but both images have slight generation issues with the clothing, and it’s unclear whether the quality-enhancing prompt actually worked.

On the other hand, at least the two generated images successfully rendered the foreground and background details, the watercolor painting style touch, and the glossy light expressions, so these three points were a success!

【In Conclusion】

I also found that trying to make clothing elaborate can cause generation issues with the clothing, so the future challenge is whether to simplify the clothing prompts or make them more specific 🤔

If anyone knows other good ways to use Seedream, please let me know 🙏

2

9 replies

A
Andrew Matthews·

nice

👍1
ね
ねむOP·

Thank you for your comment!

日
日本及大和號戰艦的应援会™·

That's a very interesting verification!

Personally, I think it might also be effective to structure the prompt by breaking it down into elements such as "character," "costume," "background," and "lighting expression." As for the costume, generating it with a simple design first, then gradually adding decorations like lace or frills and comparing the results would make it easier to understand which elements affect the output.

Also, changing color combinations, light source positions, camera angles, etc., one by one and comparing them seems intriguing. For quality improvement prompts as well, generating multiple times under as consistent conditions as possible would likely make it easier to identify trends in effectiveness.

Thank you for sharing your valuable verification results. I will use them as a reference in the future!

❤️1
ね
ねむOP·

I'm glad you said that!
Thank you so much for all the helpful suggestions—they've been really valuable! 😄

ま
まれん2·

Thank you for even incorporating my opinion!
The second one feels more flexible and has a more beautiful silhouette ✨🦋 I really love Nem-san's attitude of thoroughly enjoying the process of research image generation, thinking about what could be improved and how. I always feel that way!
I hope you never forget that feeling 👍

❤️1
ま
まれん2·

The researcher made a typo; it should have been "I researched."

ね
ねむOP·

Thank you always for your support😄 It really motivates me!
I'm also planning to do some more research on Seedream, so please look forward to it😆

🚀1
こ
こばと🕊️_ダヴ(DOVE)·

That's an interesting observation. I found it fascinating to read. I also gave it some thought myself.
image
① One of Seedream's features is that it can include text. Characters enclosed in quotation marks ("") appear accurately without distortion.
② It's easier for the prompt to be understood if written as bullet points. Since it's not a tag-based model, it's better to assume that tags are generally ineffective.
③ I've included an image and prompt as a reference. It's photorealistic, but basically, it can work for anime too, I think😂🕊️
「"Tips for using Seedream" text at the top of the screen, below it a single woman, wearing a kimono, oiran style, floral pattern, hairpin, smiling, a round window in the background with a cherry blossom garden visible, photorealistic, natural light, soft shadows」

👍1❤️1
ね
ねむOP·

Thank you for the interesting information 😄
I just looked at the other works in the gallery too! Both are output beautifully!
Next time, I'll make sure to list the prompt phrases in bullet points 👍

👍1

Leave a reply

Join the community to share your thoughts and ideas.

Sign in to join the discussion