What changed in testing (V2 vs V3, same 26 prompts and seed)

• Fuller compositions: subjects are larger and more detailed. A dragon over a castle went from a dark silhouette to a bright, detailed red dragon, and a cyclist went from small in the landscape to the main subject.

• Stronger color: boxer, samurai and storm scenes pick up magenta, violet and red light where V2 stayed darker.

• The iX look now holds on men and older people. V2 drifted toward plainer, more realistic faces on them.

• Faces are rounder with bigger eyes, so V3 reads more cartoon than V2. This is the biggest change to the style.

• Pink and magenta show up in more scenes, including ordinary ones like a kitchen or a library.

• Known characters still come from the base model. None of the gallery characters (Yoruichi, Kafka, Ranni, Boa Hancock) are in the training set, and all of them render on-model.

Tips and known limits

• Clean-shaven young men and characters with no age cues can read young, as on V2. Add stubble, a beard, mature male or mature female.

• In wide frames, use full body and say "head down to the feet". A sentence alone can crop the figure.

• Add standing if the pose matters. Some actions still crouch or sit on their own, such as feeding birds or a campfire scene.

• Neon signs can show gibberish, or even the character's name.

Recommended settings

• 30 steps, CFG 4, shift 3, euler / simple, about 1–1.3 MP (e.g. 864×1536, 1536×864, 832×1216)

• Text encoder: qwen_3_06b_base. VAE: qwen_image_vae.

• Prompt with booru tags, a plain sentence, or both.