Heirloom Male XL (SDXL) — Personal Merge, Shared

Version 2 is an upgraded refinement of Version 1, built using the same merge ingredients, but tuned for better prompt-following, stronger male/MM consistency, and cleaner results with less “drift.”
The goal stays the same: male-first photoreal, especially for MM romance/couples and fantasy/interspecies concepts — but now it’s more cooperative.

What’s improved in V2

• Better prompt adherence (less “model does its own thing”)

• MM / gay couples generate easier with normal couple prompts

• Fantasy male subjects are more consistent, including interspecies concepts when prompted

• Keeps the clean photoreal look, but with better control and reliability

🧾 Changelog (V1 → V2)

• ✅ Improved prompt-following / instruction compliance

• ✅ More consistent MM / gay couple generation

• ✅ Better fantasy male consistency

• 🎛️ Improved behavior with style LoRAs (see note below)

• 🧪 V2 sample images include an ADetailer pass (V1 samples were generated without helper tools)

📝 Prompt tips

• Best starting point: check the exact prompts used in the sample images on this model page.
  Those prompts are the closest “what this checkpoint likes” baseline.

🧠 Side note (hard-mode concepts = use guidance tools)

Some concepts are just hard mode with pure prompting — especially:
fantasy couples, interracial, interspecies, complex composition/poses.

If you use ControlNet (Canny + Reference Only) it can drastically improve consistency for:

• couple posing + composition accuracy

• identity consistency (faces/structure)

• fantasy anatomy details (tails, wings, wingspans, etc.)

🎛️ Style LoRAs + style consistency (V2 update)

Good news: in Version 2, the “style LoRA feels weaker after heavier merges” issue is largely improved — at least for the styles/LoRAs I personally use and previously had issues with.

That said, I can’t test every single LoRA or style out there. If a LoRA still feels muted, try:

If a style LoRA feels weak, try:

• Increase LoRA strength slightly

• Lower CFG a bit

• Use img2img with low denoise for a style pass

⚙️ CFG Scale tips (recommended)

Don’t crank CFG too high — it can make outputs look over-saturated / plastic.
Don’t go too low either — you’ll lose prompt control.

✅ Personal recommended range: CFG 2.5 → 6.5
(Experiment depending on prompt complexity + LoRAs)

If using IP-Adapter FaceID Plus (my V2 workflow)

• Keep CFG around ~1.5 for best results (FaceID tends to behave better at low CFG)

• If using ADetailer + IP-Adapter together:
  In ADetailer settings, set ControlNet model to “Passthrough”
  (so ADetailer doesn’t fight the identity guidance)

🚫 Suggested negative prompts (starter pack)

Use this as a baseline and adjust depending on your workflow:

Negatives:
((female, women, girl, pussy, breasts, vagina)), anime, toon \(style\), 2d, 3d, wide shot, long shot, distant subjects, painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured

🧍 If the model keeps drifting female (quick fix)

Even with a male-biased checkpoint, some prompt setups can still pull female by default — especially things like:

• “couple portrait” without specifying genders clearly

• “romance scene” + soft aesthetic keywords

• certain style LoRAs that bias feminine faces/body language

• prompts that mention “beautiful / pretty” without adding male descriptors

Try adding this tag:

• (1boy:1) with weights

  • Start without weights first (1boy)

  • If no improvement, start at 1.0 and slowly increase

  • Usually works well up to around 2.5 (don’t overdo it or it can get weird)

📸 Sample image transparency (V2)

• Same base generation settings as Version 1

• V2 samples include an ADetailer pass (mainly for faces/details)