Old Consistency V32 Lora [FLUX1.D/PDXL]

Stability checks;concept - images/tested imagesfull body - 48/48cowboy shot - 48/48portrait - 48/48close-up - 48/48**************************************<ul><li>next iteration will have a singular layer-inducing eyes pose subset, tagged with pose angles as well as including more images of eyes for each angle variation to solidify the burn. The direct eye colors will likely not be necessary, the shape however appears to be crucial beyond reproach according to my research.</li></ul>red eyes - 39/48<ul><li>full body - 6/12</li><li>cowboy shot - 9/12</li><li>portrait - 12/12</li><li>close-up - 12/12</li></ul>blue eyes - 48/48<ul><li>all poses - 12/12</li></ul>green eyes - 48/48<ul><li>all poses -12/12</li></ul>yellow eyes - 42/48<ul><li>full body - 6/12 - uncertain why this is unstable.</li></ul>aqua eyes - 48/48<ul><li>all poses</li></ul>purple eyes - 48/48<ul><li>all poses</li></ul>latex - 36/48<ul><li>close-up - 5/12</li><li>portrait - 7/12 - requires portrait and close up images</li><li>cowboy shot - 12/12</li><li>full body - 12/12</li></ul>lingerie - 36/48<ul><li>close-up - 7/12</li><li>portrait - 4/12 ? why? - requires direct portrait and close-up face images</li><li>cowboy shot - 11/12</li><li>full body - 12/12</li></ul>casual - 48/48<ul><li>all poses - 12/12</li></ul>bikini - 48/48<ul><li>all poses - 12/12</li></ul>dress - 16/48<ul><li>no pose matched the correct dress unless prompted with additional tags -&gt; need to tag dresses more specifically</li></ul>Output stability was considerably higher than expected with a plethora of usable tags already baked into pony, some including;<ul><li>&lt;color&gt; hair</li><li>&lt;color&gt; clothes</li><li>breast &lt;size&gt;</li><li>&lt;mature&gt; female</li><li>&lt;color&gt; earrings</li><li>&lt;color&gt; eyes</li><li>&lt;blurry&gt; object</li><li>&lt;area&gt; &lt;background&gt;</li></ul>There are a plethora of useful tag potentials here, so they should be played with like I did.Layering successes;<ul><li>red eyes -&gt; blue eyes; <ul><li>[red eyes:0.5], [blue eyes:0.5] -&gt; occasional bleed overlap, unstable.</li><li>red eyes, blue eyes -&gt; less consistent bleed overlap</li><li>red eyes AND blue eyes -&gt; more consistent bleed overlap, more research necessary</li><li>the majority of eyes are in fact facing the same problem, where the model imposes too much eye color over top of the eye shape, so the eye layer experiment is to be halted and implemented based on the blot experiment rather than subsequently with it.</li></ul></li><li>dress -&gt; latex<ul><li>dress, side-slit, cocktail dress, latex, latex bodysuit -&gt;forms an outfit based on multiple parts from the layers. Consistency is unstable, but the outcome is promising.</li></ul></li><li>latex -&gt; dress<ul><li>latex, latex bodysuit, dress, side-slit -&gt; forms a more dressy based object, which seems to be superimposing the dress skin gaps more than if you were to use the dress first. This implies to me the training for the dress is overfitted for the task, and thus the dress itself needs a re-evaluation.</li></ul></li><li>latex -&gt; bikini<ul><li>latex, latex bodysuit, bikini, -&gt; forms latex leggings mixed with a bikini, which also implies this is overfitted. </li></ul></li></ul>I believe I have a solution for the layering clothes overlap, and with that solution comes the solution for the eyes, skin color, and more. The majority of the consistency should come from blotting.

feminine_v11.zip

feminine_consistency_v2-loha-000500.safetensors

Files with SHA256: 70c89f5d2357817afdad5d721ce2008f0cb5ba1e114adc71b906352674f7bc85