👋 If you like what I do and want to support the development, feel free to buy me a coffee:<a target="_blank" rel="ugc" href="https://ko-fi.com/herrscheragga"><img src="https://camo.githubusercontent.com/7765e3dbce77c82639ff14d27dda26628b206727b61a6736190f90e81c78a121/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4b6f2d2d66692d537570706f72742532304d652d4646354535423f7374796c653d666f722d7468652d6261646765266c6f676f3d6b6f2d6669266c6f676f436f6c6f723d7768697465" alt="Ko-Fi" /></a><hr /><h3 id="hello-you're-probably-wondering:-why-so-many-versions">Hello, you’re probably wondering: why so many versions?</h3>Well… I’d be asking the same thing if I were in your place. The reason is simple: it’s designed this way to offer more control, since, unlike normal LoRAs, DMD2 works best at its maximum strength.For example:<ul><li>HD 1 CFG Scale has “diluted” strength, so it requires the help of triggers or manually increasing its LoRA strength. This makes it very useful for combining with PDXL LoRAs in Illustrious, since you can simply raise the strength without losing details.</li><li>DPM A1 and DPM A15 already come with boosted strength and detail, so they don’t require triggers. A1 is the standard strength, while A15 adds an extra +15%.</li><li>V4 is an experiment to generate images in 2 steps. It was created in the opposite way to HD 1 CFG: instead of reducing strength to improve stability, V4 increases strength by 1.35 ratios (20 more than DPM A15).</li><li><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/02853e8b-6148-49f0-a0ea-6a7abcc6c091/original=true/02853e8b-6148-49f0-a0ea-6a7abcc6c091.jpeg" /></li><li>V5 Enhancing checkpoint details, emphasizing styles, and making prompts more effective without changing your model. </li><li><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/38c7553a-ee46-40a1-994d-f553efedcdef/original=true/38c7553a-ee46-40a1-994d-f553efedcdef.jpeg" /></li><li>V6 It improves colors and details without changing the style; its sweet spot is cfg scale 1.</li><li>V7 (Visual Only): Built upon the foundations of V6 but heavily cleaned up to focus strictly on visual aesthetics. It sacrifices some cognitive prompt understanding (intelligence) to prioritize pure style.</li><li>V7.5 FIX (The Structural Update): The definitive, fully integrated version. It restores the deep geometric understanding and strict prompt obedience that the basic variants lacked. It features a massive parameter increase to fix the "blindness" of earlier versions, making it highly accurate for complex prompts. (Tip: Because of its dense structure, if you notice issues with hands at extremely low step counts like 3, simply increase your steps to 4-6 or lower your CFG scale to 0.6 - 1.4).</li><li><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7b4e8447-7e3f-40cf-bcd9-9576580e754c/original=true/7b4e8447-7e3f-40cf-bcd9-9576580e754c.jpeg" /></li></ul>In short: it all depends on your taste and goal. For example, V4 will produce more “noise” (details) and may sacrifice some realism unless you use it with a realistic checkpoint, V7 is great for pure stylistic overhauls, and V7.5 FIX is your heavy-duty engine for complex, highly specific prompts.<hr /><h3 id="but-what-is-this-for">But what is this for?</h3>This LoRA is based on the architecture and style of <a target="_blank" rel="ugc" href="https://huggingface.co/tianweiy/DMD2">DMD2</a>, a well-known approach for optimizing diffusion models by focusing on reducing the number of generation steps without compromising visual quality.<hr /><h3 id="so...-what-is-dmd2"><a target="_blank" rel="ugc" href="https://huggingface.co/tianweiy/DMD2">So... What is DMD2?</a></h3>DMD2 (Denoising Diffusion Probabilistic Model 2) is a variant of probabilistic diffusion models, designed to generate high-quality images from noise through an iterative denoising process. According to the literature (e.g., Ho et al., 2020, Denoising Diffusion Probabilistic Models), DMD2 optimizes the denoising process by reducing the number of steps required to achieve a quality level comparable to traditional models like DDPM.DMD2 uses an improved parameterization of the reverse diffusion process, adjusting variance weights and denoising terms to accelerate convergence. In the context of LoRAs, DMD2 serves as the base for training Low-Rank Adaptation modules that fine-tune a pretrained model (such as Stable Diffusion) for specific tasks, minimizing computational cost while preserving visual quality.In conclusion:The LoRAs described here (HD_DMD2_1_CFG-SCALE, DPM_4STEPS_A1, DPM_4STEPS_A15 and V4) are adaptations leveraging the DMD2 structure to operate with a CFG scale of 1.This is particularly interesting because normally a higher CFG scale is needed to maintain the same quality, but these LoRAs can reduce the step count to 4, 6, 8, or 10 (10 being the minimum allowed on Civitai) while achieving impressive results—cutting generation times from minutes to just a few seconds.<hr />The Compatibility Edge: The original base DMD2 is notoriously rigid—it only accepts LCM and Euler samplers and is strictly locked to a CFG scale of 1. My custom architecture breaks these limits. If you use fewer than 6 steps, it can comfortably handle up to 2 CFG. If you increase the steps to 10-14, you can push the CFG scale up to 4-5 without breaking the model, and it supports a much wider variety of sampler/scheduler combinations.<hr /><h3 id="key-features">Key Features</h3><ul><li>Optimized for fast generation: Designed to produce high-quality images with a very low number of inference steps (4, 6, or 8), enabling quick and efficient generation.</li><li>Low effective CFG scale: Works optimally around a CFG scale of 1, providing an ideal balance between creativity and fidelity without overfitting.</li><li>Three variants for different needs: Includes versions tailored for 8, 6, and 4 steps, offering flexibility depending on speed and detail requirements.</li><li>Robust visual quality: Maintains strong detail in colors, textures, and composition even with reduced steps—perfect for applications requiring both speed and quality.</li><li>Wide applicability: Suitable for users aiming to optimize generation time without sacrificing definition in their images.</li></ul><hr /><h3 id="usage-instructions-and-recommendations">Usage Instructions &amp; Recommendations</h3><ul><li>If the LoRA you’re using requires more steps to achieve a good result, you can increase the LoRA strength or add positive prompts with keywords like "hdr" to improve lighting and detail, and negative prompts like "flat color" to control saturation and shadows.</li><li>Alternatively, you can lower the LoRA strength, which allows you to use higher CFG scales without oversaturating the image. However, since this LoRA is primarily designed for CFG scale 1, the ideal strength may vary depending on your specific use case.</li><li>Experiment with both strength and CFG scale to find the optimal balance for your workflow and desired style.</li></ul><hr />Thanks so much for your support! ♥

DMD2 | Fewer Steps | SDXL | Pony | Illustrious | Noobai

This update was originally split into three variants, but I've decided to release only the definitive, most compatible version. The major breakthrough here is full integration: we've successfully recovered deep geometric understanding (<code>Attn1</code>) and merged it with strict prompt obedience (<code>Attn2</code>), entirely fixing the "blindness" of the more basic variants.Lineage &amp; Technical Leap (Intelligence vs. Weight) To fully grasp this version, it helps to look at its evolution:<ul><li>V6 (461 MB): The original baseline for "intelligence" and concept comprehension.</li><li>V7 (486 MB): Cleaned up to be purely visual, sacrificing cognitive understanding to prioritize style.</li><li>V7.5 FIX (800 MB Full / 700 MB Visual): Built on the visual foundations of V7, but massively restoring and expanding its "intelligence".</li></ul>We jumped from 254,751,648 parameters (2,784 tensors) in V7, to a massive 428,684,724 parameters (4,380 tensors) in this V7.5 FIX. It is a heavier file, yes, but it is the absolute structural solution needed to recover prompt understanding at low steps.Expectations: Low Steps and Hands I designed these iterations strictly focused on the visual aspect, without altering the knowledge base of the underlying checkpoint. Due to the pure math behind distillation models, at very low steps (e.g., 3 steps), the UNet simply doesn't have enough iteration time to sculpt perfect hands. To mitigate this, I highly recommend:<ul><li>Using strong anatomy keywords in your prompt.</li><li>Slightly increasing the step count (3 to 6 ...or +8).</li><li>Playing with lower CFG Scale values (e.g., <code>0.88-1.0</code> to <code>2.1- +4</code>).</li></ul>Training Base This version was initially forged on NoobAI and then refined on the pure foundations of the original Illustrious (I avoided community-refined mixes to keep the base style intact). Although I used the visual purity of V7 as a mold for this iteration, my technical quality standard always remains the solid foundation of V4 or V6.

Files with SHA256: ec5d578bb2976ee17f15de01dd440d570cb3726b2e09ed61f627361327813d00