oka-open-v26-lora

OKA Open v2.6 Style LoRA

Experimental style LoRA for the OH-card image workflow. It is intended to influence visible rendering behavior only: rough matte watercolor/gouache, irregular pigment shapes, broken soft edges, selective detail, and restrained facial marks.

Projection, contradiction, and psychological meaning are not trained into this LoRA. They must come from the scene prompt and human selection.

Selected Checkpoint

• File: zimage_xyqp_v26_style20_lr3e5_000000200.safetensors
• Base model: Tongyi-MAI/Z-Image
• Trigger: xyqp
• Training step: 200
• Rank / alpha: 16 / 16
• Learning rate: 3e-5
• SHA-256: b8057e1927b6333eef28a945b08c4f7e8e8c190f776cf843030f97c93e8972c9

Validated Inference Baseline

• LoRA strength: 0.25
• Steps: 28
• CFG: 4.0
• Size: 768 x 1024
• Scheduler shift: 3.0

Prompt prefix:

xyqp,

Material tail used for the accepted style samples:

medium-wide view, ordinary subdued daylight, rough matte watercolor and gouache, broad irregular pigment shapes, broken soft edges, selective detail, muted uneven colors, unprinted matte paper

Recommended negative prompt:

text, letters, words, caption, subtitle, signature, watermark, logo, card number, signage, border, frame, poster, advertisement, instructional diagram, collage, paper cutout, Japanese woodblock print, children's book illustration, photography, photorealistic, realistic skin texture, glossy, cinematic concept art, 3d render, vector art, polished commercial illustration, over-detailed, dramatic lighting, horror, grotesque, distorted hands, twisted limbs, blank face, featureless face

Scope

This checkpoint was selected because the project owner accepted the style direction in creative samples 2, 5, and 7 and then accepted the overall style of a 100-image exploratory batch. The content quality of that batch was not accepted as production-ready. Treat this LoRA as a frozen style component while prompt structure is calibrated separately.

No license grant is implied for downstream redistribution or commercial use.