Tips and tricks

To achieve strong character consistency in AUTOMATIC1111’s Stable Diffusion WebUI (especially with SDXL models like Juggernaut XL Ragnarok you’re using), you need methods that lock the face, body proportions, hair, skin tone, and overall identity across poses, angles, lighting, scenes, and partial clothing/nudity variations.

In 2026, the most reliable no-training approaches (80–95% consistency) combine IP-Adapter FaceID (for face lock) + ControlNet (for pose/body) + refinements like ADetailer. Training a custom LoRA pushes to near-100% but requires more effort.

Quick Ranking of Current Best Methods (2026 Reality)

1. IP-Adapter FaceID Plus v2 (Top recommendation for SDXL/Juggernaut) — Instant face lock from a reference image, works great for photoreal nudes/outfits.
2. ControlNet Reference / InstantID — Good for overall style + face, but FaceID is stronger for portraits.
3. ADetailer + prompt anchoring — Auto-refines faces across gens, boosts any method.
4. Custom LoRA (trained on 20–50 images of your character) — Ultimate consistency, reusable forever.
5. Fixed seed + heavy prompt description — Basic, 60–70% at best.

Step-by-Step: Best Easy Method – IP-Adapter FaceID Plus v2 (No Training Needed)

This is the go-to in 2026 for SDXL models like yours — face stays very similar even when changing poses (rear view, bent over), clothing (crop top + nude below), or scenes.

1. Download the Models (from Hugging Face):
   • Main: https://huggingface.co/h94/IP-Adapter-FaceID/tree/main → ip-adapter-faceid-plusv2_sdxl.bin (or .safetensors if available).
   • LoRA variant (stronger): ip-adapter-faceid-plusv2_sdxl_lora.safetensors (optional but recommended).
   • Place both in: F:\ai\stable-diffusion-webui\extensions\sd-webui-controlnet\models\
2. Restart WebUI (close cmd, relaunch webui-user.bat).
   • In txt2img → Scroll to ControlNet section (expand if needed).
   • In Unit 0 (or add Unit 1):
     • Check Enable.
     • Preprocessor: ip-adapter_face_id_plus (or ip-adapter-auto if it appears).
     • Model: Select ip-adapter-faceid-plusv2_sdxl (refresh dropdown with circular arrow if missing).
     • Control Weight: 0.6–0.9 (start at 0.75 — higher = stronger face lock, but can look “pasted” if too high).
     • Control Mode: “My prompt is more important” (keeps your clothing/pose prompt dominant).
   • Upload a reference face image to the “Drop Image Here” box:
     • Best: A clear, front-facing, neutral-expression photo (real or your best AI gen).
     • Crop tight to face if possible (helps lock better).
     • For your workflow: Use one good previous gen as reference.
3. Prompt Integration (keep your current structure, add face lock emphasis):textscore_9, score_8_up, score_7_up, (masterpiece, best quality, ultra detailed:1.3), photorealistic woman full body from head to toe, rear view looking back over shoulder, seductive glance directly at viewer, detailed face visible, perfect round ass focus, detailed skin pores sweat, curvy hips long legs, standing pose, bedroom soft lighting, 8k realistic anatomy, wearing only short cropped top covering small breasts, crop top only, bare midriff, completely nude from waist down, naked ass exposed, bare buttocks pussy labia visible detailed from behind, no lower clothing no panties no shorts, (small perky breasts:1.3), (petite bust:1.25)Negative (keep strong on unwanted elements):textno face visible, pants shorts skirt thong panties bikini bottom underwear covered ass covered pussy full clothing large breasts big tits huge boobs deformed face bad eyes asymmetrical eyes blurry low quality cartoon anime mosaic censoring watermark text score_6 score_5 score_4
4. Generation Settings (Tune for Consistency + Quality):
   • Resolution: 768×1152 or 896×1344 (taller for full body head-to-toe).
   • Steps: 35–50.
   • CFG: 7–8.5.
   • Sampler: DPM++ 2M Karras.
   • Hires fix: On (denoising 0.3–0.4, upscale 1.5–1.8, 4x-UltraSharp) — restores sharpness.
   • Batch size: 4–8 (variance helps find best face match).
   • Seed: Lock your good one initially, then vary slightly.
5. Refinements for Even Better Results:
   • Add second ControlNet unit: For pose lock (Unit 1: openpose_full preprocessor + openpose model, weight 0.8–1.0, upload rear/full-body pose reference).
   • ADetailer: Enable → face_yolov8n.pt or mediapipe_face → Denoising 0.35–0.5 (auto-fixes face details).
   • If face drifts: Lower FaceID weight to 0.65, or generate 10+ and pick best → use as new reference.

If You Want Near-Perfect Consistency (Training Route)

• Generate 20–50 varied images of your character using the above method.
• Train a LoRA with Kohya_ss (GUI) or WebUI built-in trainer (rank 16–32, 1000–3000 steps).
• Trigger in prompt: <lora:yourcharacter:0.7–0.9>.
• Combine with IP-Adapter for hybrid strength.

Quick Tips for Juggernaut XL

• It excels at photoreal + prompt obedience — FaceID works very well here.
• If clothing (top only) or labia visibility fights, add weights like (labia detailed exposed:1.3) in positive, (covered pussy:1.5) in negative.

Start with IP-Adapter FaceID + your reference face — generate a few full-body rear views. The face should stay locked while ass/labia/top remain as prompted. If results vary too much or face quality drops, paste a description of the issue + seed for tweaks. This setup should get you 85–95% consistency without training!