Twinx is an offshoot of <a target="_blank" rel="ugc" href="https://civarchive.com/models/628763/androflux">AndroFlux</a>. It generally makes more twinkish men than the base Flux model without much additional prompting. It won't go full twink all the time and can still produce some muscular men, but it's still a bit more narrowly focused than AndroFlux and I want to keep them separate.You can find my ComfyUI workflow here: <a rel="ugc" href="https://civarchive.com/models/696466/flux-prompt-saver">Flux Prompt Saver</a>Version 1 currently struggles with:<ul><li>Prompt adherence</li><li>Hands and other anatomy</li><li>Bulky/hairy men</li><li>Darker skin tones</li><li>Non-photographic art styles</li></ul>This LoRA can technically do NSFW as well, but it needs work.Circumcised is considered the default here so including it in your prompt will probably produce the opposite. This area is undertrained.

[Flux] Twinx

Initial version. Trained for 2000 steps at 512/768, 2600 more steps at 1024, and then a final 2100 step pass at 1024px with additional closeups. 5000 steps approx. equals 26 epochs with the initial dataset. This uses a much smaller dataset than AndroFlux and thus was limited in scope but was faster to train. Training took 3 days total on an RTX 6000 Ada.Training settings<ul><li>Learning rate: 0.0001</li><li>Effective batch size: 4<ul><li>Micro-batch size: 4</li><li>Gradient accumulation steps: 1</li></ul></li><li>Optimizer: adamw_bf16</li><li>Precision: bf16</li><li>Quantized: Yes: int8-quanto</li><li>Xformers: Not used</li><li>LoRA Rank: 64</li><li>LoRA Dropout: 0.1</li></ul>

twinx_v1.safetensors

Twinx is an offshoot of <a target="_blank" rel="ugc" href="https://civitai.com/models/628763/androflux">AndroFlux</a>. It generally makes more twinkish men than the base Flux model without much additional prompting. It won't go full twink all the time and can still produce some muscular men, but it's still a bit more narrowly focused than AndroFlux and I want to keep them separate.You can find my ComfyUI workflow here: <a rel="ugc" href="https://civitai.com/models/696466/flux-prompt-saver">Flux Prompt Saver</a>Version 1 currently struggles with:<ul><li>Prompt adherence</li><li>Hands and other anatomy</li><li>Bulky/hairy men</li><li>Darker skin tones</li><li>Non-photographic art styles</li></ul>This LoRA can technically do NSFW as well, but it needs work.Circumcised is considered the default here so including it in your prompt will probably produce the opposite. This area is undertrained.

&lt;p&gt;Initial version. Trained for 2000 steps at 512/768, 2600 more steps at 1024, and then a final 2100 step pass at 1024px with additional closeups. 5000 steps approx. equals 26 epochs with the initial dataset. &lt;br /&gt;This uses a much smaller dataset than AndroFlux and thus was limited in scope but was faster to train. Training took 3 days total on an RTX 6000 Ada.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;Training settings&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;Learning rate: 0.0001&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Effective batch size: 4&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;Micro-batch size: 4&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Gradient accumulation steps: 1&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Optimizer: adamw_bf16&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Precision: bf16&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Quantized: Yes: int8-quanto&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Xformers: Not used&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;LoRA Rank: 64&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;LoRA Dropout: 0.1&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;

Files with SHA256: a190425d26691d441a09f11c8503df53e30a6634020327a4aec2ffe6a8e3ed6b