ReaLiuM

🧪 ReaLiuM v2 — The vPred Bridge to Renaissance v5

ReaLiuM v2 is the direct preview of my training path for Illustrious Renaissance v5. The goal: migrate to vPred and measure how prompt response, color control, and high-res stability change. It’s still the house experimental line, but the quality places it between v4 and v5—cleaner, more consistent, and with tighter tonal control.

🔍 What’s new in v2

• vPred architecture (preview): more stable responses with low CFG, less color washout.

• Quality “between v4 and v5”: improved sharpness and micro-detail without losing the Renaissance look.

• Prompt adherence: handles complex scenes and short prompts very well.

• 1536p+ stability: fewer artifacts and banding, more consistent contrast.

• Negatives optional: solid “out of the box”; only add a couple if you push very dense scenes.

⚙️ Recommended (vPred)

• CFG: 1–5 (start at 3; drop to 2 if it feels stiff; raise to 4–5 if it lacks intent).

• Steps: 25–35.

• Sampler: Euler A

• Base resolution: 1152–1536 long side.

• LoRAs: 0.5–0.9 depending on intensity (keep prompts clear).

💡 Quick Tips

• Short prompts → cleaner, editorial look with tight color.

• Dense prompts → add +3 steps and keep CFG ≤3 to avoid over-contrast.

• For odd aspect ratios: lower CFG 0.5–1 or add 1–2 generic negatives (worst quality, low quality).

This is a bridge to Illustrious Renaissance v5. If it works here, it should work even better in v5.
Try it. Abuse it. Break it. Or fall in love with it. Either way… it’s alive.

— DeViLDoNia

---

---

🧪 ReaLiuM ReFoRGeD v1.01 – The Chaos Lab Got an Upgrade

What started as a failed offshoot is now a full-blown experimental platform. ReaLiuM ReFoRGeD is no longer just scraps – it's a nearly stable model built for testing wild ideas before (or instead of) integrating them into Illustrious Renaissance.

LoRAs, model merges, encoder re-trainings, deep prompts, shallow madness… this thing has it all. And somehow, it works.

🔍 What to expect in v1.01:
✅ Massively improved prompt adherence
✅ High-quality generations at 1536 resolution without breaking
✅ Great anatomy (yes, even hands… mostly)
✅ Sharper, cleaner outputs than the original ReaLiuM
🎨 From gritty realism to stylized elegance – it adapts surprisingly well.
⚠️ Some edges still rough. It’s an evolving project.

This is no longer an alpha. This is ReaLiuM with a purpose – a fusion reactor where techniques are tested, broken, and reforged into something... usable. Maybe even good.

🧠 No generation tips needed – use it like you would Renaissance, but expect some surprises.
💡 Fun Fact: I created this just to throw stuff at it. Now I use it more than I expected. That’s either progress or Stockholm syndrome.

Try it. Abuse it. Break it. Or fall in love with it.
Either way… it’s alive.

– DeViLDoNia

---

---

🧪 ReaLiuM Alpha v1.0.B55 – A flawed gem or just another discard?

Born from the scraps of Illustrious Renaissance training sessions, ReaLiuM is a strange beast. Some things it does are absolutely awful… but what it does get right, it nails better than any other realistic model I've released or seen published.

🔍 What to expect:

• ✅ Higher image quality than Illustrious Renaissance, thanks to training on higher-resolution datasets.

• ❌ Not good prompt adherence.

• ❌ Hands... yeah, they're a mess.

• ⚠️ Barely tested. Honestly, I haven’t had time to go deep with it.

I’m releasing this as an experimental Alpha, not because it's ready, but because some prompts gave me such stunning results that I hesitated before deleting it entirely.
So… let the users decide:

Should I keep training this model alongside Illustrious Renaissance, or abandon it as another failed experiment?

🚫 No generation tips this time. Just the showcase images as reference. Try it if you feel like it. Feedback will guide the future.

💡 Fun fact: Someone offered me money to release it elsewhere after seeing an image made with this alpha. I declined. I don’t do this for money.
I do it because my life is sad and I’m a forever alone. 😅

- DeViLDoNia