The quality of your LoRA is decided before training — in the images you collect and the captions you write. These recommendations walk through each EDICO step with the numbers and rules of thumb that work.

• Pick a model marked [VL] — that's EDICO flagging it as vision-capable.
• Recommended: huihui_ai/qwen3-vl-abliterated:8b-instruct — strong detail and instruction-following.
• Budget hardware? huihui_ai/qwen2.5-vl-abliterated:7b or a 4b variant run faster with slightly less nuance.
• Consistency beats cleverness — whatever you pick, finish the whole dataset with the same model so wording stays uniform.

• Use a unique, uncommon token — avoid real names or single words already common in captions (e.g. "girl"). A fictional tag like Delia450K2 is far easier for the model to treat as one concept.
• Keep it consistent — same spelling, same format, across every image and every caption.
• Write a rich, fixed base identity — hair, eyes, build, skin tone. Because it's prepended automatically, every caption carries the full character description.

• Aim for 50–200+ good images. For a single character, ~100 is a strong starting point; more helps rare angles or outfits.
• Prefer square standard resolutions — 512×512, 768×768 or 1024×1024. EDICO flags non-standard sizes in red on every card.
• Seek variety on purpose:
  • poses (standing, sitting, dynamic…)
  • framings (close-up, half body, full body…)
  • gazes, expressions, and camera angles
  • lighting and backgrounds (studio, daylight, indoors…)
• Keep the face clearly visible in most images — that's what your LoRA has to learn hardest.
• Drop duplicates and near-duplicates before you start. EDICO's DUP badge helps you spot caption twins.
• Remove watermarks, text overlays, and extreme crops — the model will treat them as features of your character.

• Balance categories — keep pose, framing, viewpoint, lighting and background terms genuinely distinct so each caption describes one of each.
• Avoid overlapping terms across categories ("sitting" in Pose and "sitting on chair" elsewhere will confuse matching). EDICO's filter pills make overlaps visible.
• Keep terms short and consistent — "looking at camera" not "gazing toward lens". The AI reproduces whatever you teach it.
• Clothing & accessories are free-text because outfits are endless — but keep the wording style parallel with category terms.
• Test one small batch first — caption 5–10 images, review in the Organizer, then tune the vocabulary before running the full set.

• Quality fluff — "masterpiece", "8k", "photorealistic". These train the LoRA to chase styles, not your character.
• Articles & full sentences — "a woman wearing a jacket" adds noise. Keyword style is cleaner.
• Emotion guesses — describe the visible expression, never inner feelings.
• Identity drift — if the same character's hair changes color between images, your LoRA won't know which one to learn.
• Mixed languages — keep captions in one language throughout.

• Scan red resolution badges — re-export or drop non-standard sizes.
• Review DUP badges — confirm they're intentional variety, not duplicates.
• Filter by unspec in each category — those are images the AI couldn't categorize; fix them by hand in the editor.
• Rearrange deliberately — group close-ups first, or order by outfit — then run the sequencer so your final folder is clean and numbered.
• Grab the folder with the Open files button — /working now holds your numbered .png/.txt pairs, ready for your trainer (Kohya, OneTrainer, etc.).

📚 More resources