<h1 id="frankenstein-model-suite:-experimental-grounds-j3lqj9z9v"><strong><span style="color:#82c91e">Frankenstein</span> - Model Suite: Experimental Grounds</strong></h1><ul><li><p><strong><span style="color:#fa5252">Warning</span>: <span style="color:#868e96">Highly Experimental Models Ahead! Proceed with Curiosity (and Caution)!</span></strong><br /></p></li></ul><p>Welcome to the laboratory! The Frankenstein Model Suite is where I unleash my more... <em>unconventional</em> creations. Unlike my refined models such as Everlasting, the models hosted here are born from experimental merges, unusual training techniques, bespoke modification scripts, and sometimes, sheer chaotic curiosity.<br /></p><p>Think of this as a collection of works-in-progress, oddities, and boundary-pushing attempts. They are stitched together using various methods, potentially including techniques developed with my custom-crafted merger tool and other experimental approaches</p><p></p><h1 id="methods-swyyt7bg3"><br />Methods</h1><h3 id="block-swap-(bs-custom-merger-implementation):-instead-of-blending-weights-this-method-swaps-entire-architectural-blocks-(like-input-middle-or-output-sections-of-the-u-net)-between-models-based-on-certain-rules.-it-aims-to-combine-functional-strengths-but-risks-incompatibility-between-the-stitched-sections.-(highly-experimental-but-functional)-dqdshklhw"><em> </em><span style="color:#fab005">Block Swap (BS - Custom Merger Implementation):</span> Instead of blending weights, this method swaps entire architectural blocks (like input, middle, or output sections of the U-Net) between models based on certain rules. It aims to combine functional strengths but risks incompatibility between the stitched sections. <em>(HIGHLY EXPERIMENTAL BUT FUNCTIONAL)</em></h3><h3 id="latent-resonance-tuning-(lrt):-attempts-to-identify-key-latent-features-in-one-model-and-subtly-amplify-corresponding-weights-in-another-trying-to-make-the-second-model-&quot;resonate&quot;-with-the-first-model's-specific-strengths-at-a-feature-level.-(experimental)-hxf26w1r6"><em> </em><span style="color:#82c91e">Latent Resonance Tuning (LRT):</span> Attempts to identify key latent features in one model and subtly amplify corresponding weights in another, trying to make the second model "resonate" with the first model's specific strengths at a feature level. <em>(EXPERIMENTAL)</em></h3><h3 id="gradient-harmonic-injection-(ghi):-involves-injecting-scaled-gradients-from-a-parent-model-during-a-post-merge-fine-tuning-phase-attempting-to-subtly-steer-the-merged-model-towards-a-parent's-style-without-full-retraining.-(experimental)-hneb20nox"><em> </em><span style="color:#12b886">Gradient Harmonic Injection (GHI):</span> Involves injecting scaled gradients from a parent model during a post-merge fine-tuning phase, attempting to subtly steer the merged model towards a parent's style without full retraining. <em>(EXPERIMENTAL)</em></h3><h3 id="cross-attention-splicing-(cas):-mixes-components-of-the-cross-attention-mechanism-(e.g.-querykey-matrices-from-one-model-valueprojection-from-another)-to-blend-how-models-interpret-prompts-with-how-they-apply-that-interpretation.-(experimental)-cmfg4foda"><em> </em><span style="color:#40c057">Cross-Attention Splicing (CAS):</span> Mixes components of the cross-attention mechanism (e.g., query/key matrices from one model, value/projection from another) to blend how models interpret prompts with how they apply that interpretation. <em>(EXPERIMENTAL)</em></h3><h3 id="denoising-path-scaffolding-(dps):-uses-parts-of-a-parent-model's-u-net-to-temporarily-guide-or-replace-calculations-at-specific-steps-within-the-merged-model's-denoising-process.-(experimental)-td3gktpm7"><em> </em><span style="color:#15aabf">Denoising Path Scaffolding (DPS):</span> Uses parts of a parent model's U-Net to temporarily guide or replace calculations at specific steps within the merged model's denoising process. <em>(EXPERIMENTAL)</em></h3><p></p><p></p><h1 id="what-to-expect:-aa9c62ryr"><strong>What to Expect:</strong></h1><h3 id="unpredictability:-results-can-vary-wildly.-what-works-for-one-prompt-might-break-on-another.-expect-the-unexpected!-9b8klk2ay"><em> </em><span style="color:#e64980">Unpredictability:</span> Results can vary wildly. What works for one prompt might break on another. Expect the unexpected!</h3><h3 id="potential-instability:-some-models-might-produce-artifacts-glitches-or-struggle-with-coherence-more-than-polished-releases.-hands-feet-and-eyes-might-be-even-more-challenging-than-usual.-ppiizz4th"><em> </em><span style="color:#fd7e14">Potential Instability:</span> Some models might produce artifacts, glitches, or struggle with coherence more than polished releases. Hands, feet, and eyes might be even more challenging than usual.</h3><h3 id="niche-capabilities:-a-model-here-might-do-one-very-specific-thing-interestingly-but-be-terrible-at-everything-else.-qc8c4ajj6"><em> </em><span style="color:#fab005">Niche Capabilities:</span> A model here might do one <em>very specific</em> thing interestingly, but be terrible at everything else.</h3><h3 id="occasional-gems:-sometimes-the-chaos-yields-surprisingly-unique-aesthetics-or-combinations-you-won't-find-elsewhere!-4abne7kz9"><em> </em><span style="color:#4c6ef5">Occasional Gems:</span> Sometimes, the chaos yields surprisingly unique aesthetics or combinations you won't find elsewhere!</h3><p></p><p></p><h1 id="why-this-suite-exists:-qerre5t5d"><strong>Why This Suite Exists:</strong></h1><p>This is a space for open experimentation and sharing explorations that aren't quite ready for a "stable" release. It's a chance to play with rougher ideas and see what strange beauty (or monstrosity) might emerge. Your feedback and discoveries – good, bad, or just plain weird – are part of the experiment!</p><p></p><p></p><h2 id="recommended-settings:-iu6bj7ndc"><strong>Recommended Settings:</strong></h2><h3 id="highly-model-dependent!-there-are-no-universally-recommended-settings-here.-check-individual-model-notes-if-available.-cbik6jnd0"><em> </em><span style="color:#e64980">Highly Model-Dependent!</span> There are NO universally recommended settings here. Check individual model notes if available.</h3><h3 id="starting-point:-you-can-start-with-common-settings-(e.g.-sampler:-euler-a-steps:-20-30-cfg-scale:-5-7-clip-skip:-1-2)-but-be-prepared-to-experiment-heavily.-vemzhvrcx"><em> </em><span style="color:#15aabf">Starting Point:</span> You <em>can</em> start with common settings (e.g., <code>Sampler: Euler a</code>, <code>Steps: 20-30</code>, <code>CFG Scale: 5-7</code>, <code>Clip Skip: 1-2</code>), but <strong>be prepared to experiment heavily.</strong></h3><h3 id="prompting:-some-models-might-require-very-specific-trigger-words-or-prompt-structures.-others-might-ignore-negative-prompts-entirely-or-react-strangely-to-them.-pay-attention-to-any-included-examples.-luk5yakd6"><em> </em><span style="color:#82c91e">Prompting:</span> Some models might require very specific trigger words or prompt structures. Others might ignore negative prompts entirely or react strangely to them. Pay attention to any included examples.</h3><h3 id="denoising:-for-img2img-or-refining-outputs-denoising-strength-might-need-careful-tuning-(try-ranges-like-0.5-0.75).-3ro7zo5xa"><span style="color:#7950f2"> Denoising:</span> For img2img or refining outputs, denoising strength might need careful tuning (try ranges like 0.5-0.75).</h3><p></p><p>Usage Guidelines:</p><p> *Non-commercial Use:** Primarily intended for personal experimentation and non-commercial creative exploration. Please generate responsibly.</p><p> *Sharing:** Feel free to share your generations. If sharing the models themselves, please link back here and keep the experimental nature crystal clear.</p><p> *Attribution:** When sharing results, mentioning the specific model name and technique (if known) from the Frankenstein Suite helps track these experiments.</p><p> *Feedback Welcome:** If you find interesting settings, prompts, or use cases (or run into spectacular failures), sharing your findings can be valuable!</p><p></p><pre><code>Dive in, experiment (safely!), and let's see what these stitched-together creations can do!</code></pre><p></p>

Frankenstein - Model Suite

<p>Created by streamlining two base models and selectively removing blocks from models known for superior hand rendering, these components were then merged. The resulting model was also enhanced with a custom LoRA merge, employing experimental culling and cropping training techniques to improve hand generation by 38%, while minimally impacting the overall style by just 5%.<br /><br />Used by combining the trigger words to improve hands, without using the triggers hand improvement is about 25%</p>

Files with SHA256: 9cfd563b67beb797b34ed803e35f19e6f75716e4cc6dde2b2b930c659abb3790