🧠 DMD2: Smart-Brain Infusion — Portable Checkpoints & Pure Style Modular System

👋 If you like what I do and want to support the development, feel free to buy me a coffee:

I'm back, and this time I'm bringing my most advanced, flexible, and powerful system yet!

Since my last milestone release back on June 9, 2026 (based on Better Days v0.2), I’ve been quietly grinding behind the scenes. Driven by personal R&D and a few great discussions in DMs regarding Anima compatibility, I decided to test its limits and explored two distinct development paths:

• Extracting Anima and compiling it to be 100% compatible across SDXL.

• Engineering a hyper-optimized version built exclusively for the Anima environment.

Because these Anima-specific builds still require further testing and benchmarking before a public release, I’m keeping them in the lab a bit longer. For now, I decided to focus my full attention on perfecting and releasing the core modular system and files currently in my hands—ensuring you get maximum stability right out of the box.

While those previous builds did a fantastic job enhancing Velvette and other similar checkpoints, I felt something was still missing: that "recognizable spark"—the raw, unadulterated essence of the base render style.

However, locking down that exact style was no easy task—a single minor shift in tensor weights and the whole thing instantly becomes incompatible. That was the real battle, of course...

As a huge bonus discovered during these hard-fought experiments, this extraction framework opens a goldmine for nostalgia lovers: bringing back classic SD 1.5 styles into the modern era.

SD 1.5 holds years of legendary, community-crafted checkpoint aesthetics that were unfortunately lost or washed out during the industry transition to SDXL. With this new modular extraction methodology, we can now:

• Mine vintage, handcrafted render signatures and unique brushstrokes.

• Isolate pure visual DNA without carrying over old 512x512 resolution limitations or outdated anatomical flaws.

• Port them directly onto high-resolution SDXL, Illustrious, and NoobAI bases.

You get the iconic, nostalgic artistic flair of classic 1.5 models powered by modern 1024px+ geometry and prompt precision!

Why One File Wasn't Enough: The Modular Solution

For months, I tried to force a single model to learn both complex anatomical logic and heavy rendering traits simultaneously. The result? A constant tug-of-war in the tensors that compromised either the structure or the texture.

After endless trial and error, deep tensor research, and late-night Colab sessions, I hit the breakthrough. The secret wasn't forcing everything into one file—it was separating the Brain from the Style.

Instead of a monolithic extraction, I evolved this workflow into a 3-Step Modular Pipeline:

1. Concept & Brain Extraction: Isolating pure anatomy, spatial awareness, and text-encoder logic (associating prompts directly with concepts).

2. Pure Base Style Extraction (100% Faithful): Isolating the exact visual signature—brushstrokes, lighting, color palette, and render weight—directly from the source checkpoint without carrying unwanted artifacts.

3. Orthogonal Concatenation & Fusion: A merging methodology that allows both components to run concurrently without competing or overbaking the image.

The Fused vs. Separated Reality

Let’s talk about Fusions vs. Separated LoRAs.

Is a single-file fusion viable? Yes, absolutely. In fact, fused files bring incredible advantages: enhanced micro-details, superior definition, and 1-click convenience without loading multiple LoRAs.

However, tensor physics comes with a trade-off: fusing DMD2 with Style slightly alters the base aesthetic. Instead of a 100% raw, flat cell-shaded render, a fused file tends to lean toward a more polished, 2.5D volumetric look. Furthermore, fused files are mathematically heavier.

How We Are Distributing This:

To guarantee maximum visual accuracy and full modular freedom, we are prioritizing standalone, individual files over monolithic merges:

• 🟢 Standard Release (Individual Separated Files): The Brain and Style are now uploaded as separate, individual .safetensors files in the download list. Loading both in your WebUI (<lora:Brain:1.0> <lora:Style:1.0>) is the recommended gold standard to achieve a 1:1 exact visual match to the original checkpoint.

• ⚡ Direct Style Access: For your convenience, the matching Style LoRA is attached right here alongside the Mini Brain for quick 1-click downloading. You can also find it indexed in our dedicated Style Section repository for use with other models.

• 🟡 Fused / Hybrid Builds: Single-file fused builds are no longer our primary release focus. However, they remain available under the "Other Files" download section as optional extras for casual 1-click workflows or heavy detail boosting.

Research Status: DMD2 + Fusion development is still actively ongoing. The results are extremely promising, but because DMD2 is so potent, I hit a mathematical "signal wall." The current path forward involves reducing specific UNet layers to balance the signal. It's a process of trial and error, and I'm not throwing in the towel anytime soon! For now, separated files remain our gold standard for absolute fidelity.

🎨 New: Dedicated "Style Section"

We are expanding! Pure style extraction deserves its own focused ecosystem, so I am introducing a dedicated Style Section:

• Reclaim & Port Unique Styles: This section is designed to recover aesthetic signatures from obscure LoRAs or checkpoints that only worked on specific base models, allowing you to port them seamlessly into SDXL/Illustrious/NoobAI.

• Rank Optimization (R32 to R256): I am actively testing different rank compressions (from Rank 32 up to Rank 256) to drastically shrink file sizes while keeping visual loss to an absolute minimum.

• The Ultimate Duo: The Style Section acts as the direct counterpart to our Portable Checkpoints/Brains. While style files work great standalone, pairing them with a Brain LoRA creates an unbeatable dynamic duo.

• Fast-Access .zip Bundle: To save you time, if you are grabbing a Mini Checkpoint from this post, the matching standalone Style file is included right in the .zip download!

The LoRA Types Explained: TOTAL, DUPLO & ULTRA

Okay, now let's talk about these Loras, These LoRAs function as "Structural Converters."

Think of them as a high-end "Cosplay" for your checkpoints: they allow a lightweight model to adopt the exact visual DNA, intelligence, and stylistic precision of a massive 6GB checkpoint (like Pony or Illustrious).

Instead of dealing with architectural instability or NaNs, I have distilled the core features of these giant models into optimized 500MB-1GB files. They let you inject the prompt-understanding and "soul" of a heavy base model into any other refined checkpoint without the overhead of downloading or loading 6GB files every time.

• In conclusion: Now you have a tool for every need:

  • Do you just want a refined DMD? -> Total.

  • Do you want the information and style? -> Duplo.

  • Do you want it all? It'll be a Copy -> Ultra.

🟢 TOTAL (Concept Injector):

• What it extracts?: Text Encoders + attn2 (Cross-Attention).

  • Complete Version: Extracts Text Encoders + attn2 (Cross-Attention). It’s the "Brain" of the model.

  • Visual Only Version: Extracts only the Style (UNet). | DMD2 Pure.

• What does it do?: Associates words with concepts.

  • Total knows that "Miku" means "Teal hair, long pigtails".

• Result: Corrects what is drawn, but the "brushstroke" remains from your base model.

🟡 DUPLO (Structure & Geometry):

• What it extracts?: Text Encoders + attn2 + attn1 (Self-Attention).

• What does it do?: Controls geometry and spatial composition. attn1 is where the "shape" of the style resides (eye size, body proportions, composition).

• Result: The image gets the structure of the source model (e.g., Pony), but the rendering (skin, lighting) is a hybrid.

  • Best for fixing anatomy while keeping your checkpoint's texture.

🔴 ULTRA (Full Replica):

• What it extracts?: EVERYTHING (attn, ff, proj, te).

• What does it do?: Copies the FeedForward (FF) layers too, which determine the Render Style (lighting, line weight, shading).

• Result: A complete conversion. The base model visually disappears and becomes a perfect replica of the source.

IMPORTANT VERSIONS & WARNINGS

⚠️ Visual Only (Total) vs. Complete (Zip)

• Visual (Online Gen Friendly): Use this for quick style transfer.

• Visual Ultra: Use this for quick checkpoint style transfer.

• Complete (Zip): Includes the "Fixed" files that connect text properly. Use this for serious work.

💡 Visual Only Usage Note: Don't be scared! Even though this is based on my DMD2 architecture:

• It works perfectly at HIGH STEPS (20-30+) without burning (great for detailing).

• It works perfectly at LOW STEPS (4-8) for speed. Wink wink! 😉

🛡️ BONUS / EXTRA: Neural Repair & Runtime Fixer

(Originally planned as a sampler update, but shifted to fix a massive community-wide flaw!)

Many popular merged checkpoints in the wild suffer from a corrupted Layer 11, causing:

• ❌ Text Encoder errors and NaNs (black images/crashes).

• ❌ Poor LoRA compatibility.

• ❌ Massive information/tensor loss.

• And here's the solution: [Anti-Nans + RAM Cleaner] Uh-huh... The LoRA repair method failed, so I engineered a Runtime Fixer. Just paste the provided script into a new cell in your Colab/Notebook, run it before the WebUI, and you are good to go.

  • Herrscher Shield: Scans Layer 11 and eliminates NaNs in RAM instantly. No need to download fixed checkpoints.

  • AGGA Optimizer: Aggressively cleans RAM to prevent Colab crashes.

🎁 Christmas Release Announcement: The Fusion & Colab Scripts

I know many of you are eager to get your hands on the internal workflow.

Months of hard work went into perfecting these Colab scripts, the Anchor-filtering mechanics (bringing Pony compatibility to Illustrious/NoobAI cleanly), and the 3-step extraction pipeline.

• Community Support Matters: Likes and interactions are what keep these projects visible and funded on public platforms.

• Direct Messages (DM): If you are interested in trying specific scripts, send me a DM! While I won't give away the core master scripts all at once (I might share one of the three key scripts via DM or explain parts in my upcoming articles), I want to reward those who have supported this journey from the very beginning.

• Global Release: On December 24th (Christmas Eve) 🎅🎄, I will publicly release the complete script suite and Colab notebook for everyone!

Thank you so much to everyone who tests, likes, and supports this project. Enjoy the new Mini Checkpoints, grab the style files, and get ready for the Christmas release! ♥