Guide — Using your LoRA: “painterly semi-realistic anime” with SDXL in PixAI

Nice — here’s a clear, practical, copy-pasteable guide that shows exactly how to use your painterly LoRA with the SDXL model in PixAI. I kept it focused and tested against the common SDXL/LoRA behavior so it’s ready to follow step-by-step.

Assumption: you’ve already trained/exported a LoRA file (e.g. haerin_painterly_v1.safetensors) for SDXL (recommended). If your LoRA was trained on SD 1.x it may not transfer perfectly — see Troubleshooting below.

---

1) Quick checklist (what you need)

• SDXL model available in PixAI (select SDXL in the model dropdown).
• The LoRA file exported (e.g. haerin_painterly_v1.safetensors) and uploaded/installed into PixAI’s LoRA/weights folder or via PixAI’s UI import.
• A subject prompt to prepend the style token to (examples below).
• Basic PixAI settings: sampler, steps, cfg/guidance scale.

---

2) How to load the LoRA (PixAI — typical steps)

1. Open PixAI and pick the SDXL model (e.g., “sdxl” / “SDXL-v1”).
2. Locate LoRA / Model Weights area in PixAI (often labeled LoRA, Custom weights, or Extensions).
3. Upload your .safetensors LoRA file or place the file into PixAI’s expected LoRA folder and click Load/Refresh.
4. Confirm the LoRA appears in the UI (name shown in the LoRA list). If PixAI offers a quick-enable checkbox, enable it.

If PixAI version you use doesn’t accept raw LoRA files, you’ll need a front-end (Automatic1111 / ComfyUI / local SDXL) to test the LoRA first or convert it to a format PixAI accepts.

---

3) Prompt format & where to place the LoRA token

Place the LoRA token at the start of the prompt (this is the safest placement for SDXL to ensure the style conditions the generation early):

<lora:haerin_painterly_v1:0.8> [subject description here], painterly semi-realistic anime style, (optional extra style keywords)

• haerin_painterly_v1 — replace with your LoRA filename (no extension) or canonical name shown in PixAI.
• 0.8 — LoRA weight (recommended starting point). Range 0.0–1.0 (can exceed 1.0 in some UIs but start ≤1.0).

---

4) Example prompts (ready to paste)

Portrait (close-up)

<lora:haerin_painterly_v1:0.8> close-up portrait of a young woman, 3/4 view, short tousled black hair, soft smile, delicate eyes, soft painterly brushes, warm rim light, cinematic chiaroscuro --ar 2:3

Full-body / character

<lora:haerin_painterly_v1:0.85> full-body portrait of a schoolgirl walking, dynamic pose, loose painterly clothing texture, soft sunlight, visible impasto brush strokes, cozy atmosphere --ar 9:16

Scene / environment

<lora:haerin_painterly_v1:0.75> two girls by a hedge at golden hour, impressionistic brushwork, warm ambient light, loose background, subtle canvas texture --ar 3:4

Using an image reference (img2img)

• Load reference image. In prompt field:

<lora:haerin_painterly_v1:0.9> painted reinterpretation of the provided photo, painterly semi-realistic anime, soft brush strokes, warm rim light

• Denoising / Strength: 0.2–0.5 for a faithful painterly conversion.

---

5) Recommended SDXL & PixAI settings

• Model: SDXL (select the SDXL variant available).
• Sampler: DPM++ 2M Karras or Euler a (both work well).
• Sampling steps: 20–40 (30 is a good default).
• Guidance / CFG scale: 6.0–9.0 (start at 7.0).
• Resolution: Portrait: 1024×1536 or 1536×2048 for higher detail. Full-body: 768×1280 or 1024×1536. (Match SDXL VRAM capacity.)
• Batch size: 1 (for best preview iteration).
• Seed: leave random for variety or set for reproducible results.
• LoRA weight: 0.6–0.95 typical. Start 0.8 and tune ±0.1.
• Negative prompt: photorealistic, lineart, pixel art, 3d render, watermark, text, oversharpened, noisy artifacts

---

6) Tips to get the painterly look without losing subject

• Combine tokens: Put precise subject tokens first, LoRA token first or immediately before subject depending on PixAI. Example: <lora:haerin_painterly_v1:0.8> young woman, 3/4 portrait,...
• Add texture words to the prompt to push painterliness: impasto, canvas texture, visible brush strokes, soft blended edges.
• If faces are too stylized or lose identity: lower LoRA weight to 0.5–0.65 and explicitly add detailed realistic eyes, natural face anatomy.
• If style is too weak: increase LoRA weight by +0.05 increments or add painterly keywords.
• For consistent series/character art: fix the seed and keep LoRA weight constant; optionally lock face descriptors (hair, eye color).

---

7) Img2Img workflow (best way to turn a photo into painterly art)

1. Choose SDXL model. Load source image.
2. Prompt: <lora:haerin_painterly_v1:0.9> painted reinterpretation of the image, soft brush strokes, warm rim light, delicate eyes
3. Set Denoising / Strength to 0.22–0.40 (lower = closer to original, higher = more stylized).
4. Use 20–30 steps, guidance 7–9.
5. Review and tweak LoRA weight and denoising until you like the painterly transform.

---

8) Combining LoRAs & layering

• You can combine LoRAs: <lora:haerin_painterly_v1:0.7><lora:another_style:0.2>.
• Balance weights so one style stays dominant.
• If combining with a content LoRA (e.g., an artist-specific LoRA), use lower weights for additive color/texture LoRAs.

---

9) Troubleshooting & SDXL compatibility notes

• LoRA has no effect:

  • Ensure the LoRA was trained for SDXL or converted for SDXL. LoRAs trained on SD1.5 may not map 1:1 to SDXL.
  • Confirm the LoRA is uploaded and enabled in PixAI.
  • Try a higher weight (0.9) to confirm effect; if still no change, test the LoRA in a local SDXL UI (Automatic1111 / ComfyUI) to verify the file works.

• Faces / anatomy look wrong: lower LoRA weight or add corrective descriptors (realistic eyes, accurate facial anatomy). If training is available, retrain with more face examples.

• Overly painterly / too noisy: reduce LoRA weight, reduce network rank (if retraining), or remove impasto tokens.

• Style only on backgrounds / not on faces: try moving LoRA token placement to the very front of the prompt, or increase weight slightly.

• If you must convert a SD1.x LoRA for SDXL: conversion is possible but fragile — the most reliable route is to retrain or fine-tune a small LoRA on SDXL using representative images.

---

10) Quick example full run (copy/paste)

• Model: SDXL
• Prompt:

<lora:haerin_painterly_v1:0.8> close-up portrait of a teenage girl, soft tousled black hair, warm rim light, delicate eyes, painterly semi-realistic anime, visible brush strokes, canvas texture --ar 2:3

• Negative prompt: photorealistic, 3d render, watermark, text, lineart
• Sampler: DPM++ 2M Karras
• Steps: 30
• Guidance: 7.5
• Resolution: 1024×1536
• LoRA weight: 0.8

---

11) Attribution & permissions (short)

When sharing images generated with your LoRA, include a credit line:

Style: haerin_painterly (LoRA) — credit @<your_handle>

If you distributed the LoRA under non-commercial terms, remind users of that license when you share.

---

12) Advanced: training/fine-tuning notes for SDXL (brief)

• Train the LoRA directly on SDXL for best fidelity. SDXL conditioning and tokenization differ from SD1.x so style transfers are stronger when trained on the same family.
• Use network_dim 8–16 for subtler to moderate style strength. Train with consistent captions like painterly anime, soft brush strokes, warm rim light.
• Export final .safetensors and test in SDXL generation.

---