3 modi per migliorare la generazione di immagini in stile anime di Seedream

The previous article is .

In the previous article, I succeeded in verifying the hypothesis for creating beautiful anime-style images with Seedream, but the problem of poor generation of clothing descriptions remained.

So, I received various advice in the comments, and based on that, I’m trying again to create beautiful images with Seedream 👍
(Thank you to everyone who commented!)

【3 Improvement Strategies】

First, why did the poor clothing generation occur? I hypothesized that the cause was how the prompt was written and focused on improving that.

Based on the suggestions from "日本及大和號戰艦的应援会™":
"For clothing as well, first generate with a simple design, then gradually add decorations like lace and frills to compare"
and from "こばと🕊️_ダヴ(DOVE)":
"Prompts are easier to understand if written in bullet points"
I also thought of my own improvements and decided to try the following three methods:

① Generate with a simple prompt that does not instruct clothing decorations (reducing information) ⬇️
② Generate with a prompt rewritten in bullet point format (changing format) ✂️
③ Generate with a prompt that adds elements to concretize the clothing shape (adding information) ⬆️

By the way, ③ is an application of the tip I previously wrote about using Image V2: "stabilize the composition by adding prominent objects like hats" ⚖️👒
(Reference .)

This time, I experimented by modifying the part of the previous prompt: "a blue ribbon tie, a blue and black gothic lolita-style dress with lace and frills on the collar, hem, and sleeves"!

(Note: Again, I should add that since Seedream cannot fix the seed value, precise controlled experiments are impossible 🧪)

【Experiment Details】

① Generated with the prompt: "anime style, watercolor painting style, girl, blue eyes with light particles, black hair shining light blue, blue and black butterfly-shaped hair ornament, blue and black gothic lolita-style dress, grassland, white and pink cosmos flowers in the foreground, glowing autumn leaves trees in the background, emerald and light blue gradient sky, sparkling, lens flare, rim light."

② Generated with the prompt: "anime style, watercolor painting style, girl, blue eyes with light particles, black hair shining light blue, blue and black butterfly-shaped hair ornament, blue ribbon tie, lace, frills, blue and black gothic lolita-style dress, grassland, white and pink cosmos flowers in the foreground, glowing autumn leaves trees in the background, emerald and light blue gradient sky, sparkling, lens flare, rim light."

③ Generated with the prompt: "anime style, watercolor painting style, girl, blue eyes with light particles, black hair shining light blue, blue and black butterfly-shaped hair ornament, blue ribbon tie, blue and black gothic lolita-style dress with lace and frills on the collar, hem, and sleeves, black ribbon around the waist, grassland, white and pink cosmos flowers in the foreground, glowing autumn leaves trees in the background, emerald and light blue gradient sky, sparkling, lens flare, rim light."

※As before, due to capacity, compressed versions are shown

Previous image (before improvement) ↓

The clothing looks a bit baggy 😅

Method ① ↓

Compared to before, it became simpler and more coherent! Even just this, the details are well rendered 🔍️

Method ② ↓

Compared to before, the lace and frills around the collar are generated without issues! Also, the shape of the ribbon tie is more organized than before 🎀

Method ③ ↓

Although there is some unnaturalness in the front part of the skirt, it’s not as noticeable as the poor generation before, and the overall silhouette is more refined 🌟

So, all three were successful!

【In Conclusion】

From the above results, to generate clothing without issues:

① First, try generating with a simple prompt,
② If you want to include many elements, write the prompt in detailed bullet points,
③ Also add elements to stabilize the composition and silhouette.

With these, it seems the chances of success increase!

Also, the suggestion from "日本及大和號戰艦的应援会™" that I couldn’t try this time: "change and compare things like color combinations, light source position, and camera angle one by one" also sounds interesting and would be useful beyond Seedream!
(Considering cost, it might be safer to first try with a free model ☺️)

By the way, which image do you all prefer? Personally, I like image ② with its cute smile and image ③ which is more luxurious 😆