y2ksnap

⚙️ Training details

• Base: Krea 2 (on-site trainer, AI Toolkit engine)
• 44 curated images — mix of single-subject, couples/groups, and empty-scene shots (indoor/outdoor, day/night)
• 2,000 steps, 10 checkpoints — this upload is epoch 10
• Captions: content-only tags (subject/setting/pose) — no lighting or style words in the training captions, so the flash aesthetic stays bound to the trigger and doesn't fight your own prompt wording

🎯 Best for: night and indoor candid shots — portraits, couples, groups, and empty interiors/street scenes. Works across a wide range of subjects since the dataset deliberately spanned solo/group, indoor/outdoor, and multiple lighting times.

💡 Prompting tips

• Keep prompts short and content-only (subject + setting + time of day) — let the LoRA carry the lighting, don't re-describe flash/grain/exposure yourself
• Recommended strength: 0.8–1.2, default 1.0
• Use direct age/gender wording — "woman", "man", "young woman" etc. Avoid bare "girl" / "boy" / "teenager" without an adult qualifier; in testing, those bare words nudged renders toward a younger-looking subject regardless of intent. Same rule applies whether you're prompting adults or just describing a scene.

🚫 Not for: soft/studio lighting looks — this LoRA pushes hard toward direct-flash harshness. Lower the weight (0.6–0.8) if you want a softer version of the effect instead of full harsh-flash.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 📋 License Generated images: free to use, sell, no credit required. Model itself: not for resale/re-hosting on other generation services. Full permissions on the model page. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━