<h1 id="we're-back-baby">We're Back Baby</h1>After a two hear hiatus, we're back with a brand new Anteros model. Trained on the same imageset used to train the original and highly acclaimed <a rel="ugc" href="(https://civarchive.com/models/430454/anteros-xxxl">Anteros XXXL</a> of 1.3M professional and amateur pornography, with new high quality generated captions. This model does softcore, hardcore, and everything in-between.<h1 id="which-model-should-i-use">Which Model Should I Use?</h1><h2 id="anteros-xxx-krea2-turbo">Anteros XXX Krea2 Turbo</h2>This is the model for most users. Trained using the same Trajectory Distribution Matching used by Krea to train their turbo model, we find this model strikes a good balance between performance and flexibility. We think you'll love it.<h3 id="settings">Settings</h3>Sampler: Euler Scheduler: Simple Steps: 8 CFG: 1.0 <h2 id="anteros-xxx-krea2-raw">Anteros XXX Krea2 Raw</h2>If you find you can't quite get the result you want from the Turbo model, and are okay spending more time on generation, give this a try. It's the result of our post-finetune alignment and should have more flexibility than Turbo. This is also a good starting point if you want to train your own character LoRAs.<h3 id="setttings">Setttings</h3>Sampler: Euler Scheduler: beta57 Steps: 40 CFG: 4.5<h2 id="anteros-xxx-krea2-unaligned">Anteros XXX Krea2 Unaligned</h2>This is the raw finetune. The images it generates tend to be less visually appealing. The only reason it's here is for trainers or mergers who want to do something big with it. If you don't know why you need this, you don't.<h3 id="setttings">Setttings</h3>Sampler: Euler Scheduler: beta57 Steps: 40 CFG: 4.5<h3 id="caption-style">Caption Style</h3>Long detailed descriptions are best. Detail everything. Here are some sample captions from the actual training set:<ul><li>A naked blonde woman is posing outdoors next to a tree. She is crouching and leaning forward, looking directly at the camera with a slight smile. Her long blonde hair hangs down her right side. Her breasts are fully exposed, with her left nipple clearly visible. Her right arm rests across her body, and she is wearing orange beaded bracelets on her wrist. Her vulva is exposed and visible between her spread thighs. The background consists of blurred green and orange autumn foliage.</li><li>A woman with dark hair lies on her back with her eyes closed while being tittyfucked by a man. Her skin is glistening with sweat or lubricant. She uses her hands to squeeze her large breasts together, trapping the man's large, erect penis within her cleavage. The head of his penis is deeply inserted between her breasts, while the shaft extends outwards to the right. Her hands press firmly against her breasts to maintain the tight enclosure around his genitals. The woman's nipples are visible on her firm, rounded breasts as they envelop the penis.</li><li>A woman with dark hair and glasses is giving a blowjob to a man. She is positioned on a white padded table, leaning forward with her buttocks exposed and pushed up, wearing white stockings and black and white sneakers. Her mouth is wrapped around his large erect penis. The image is from the man's point of view, showing his chest and legs. The scene takes place in a room with a grey carpet, a blue wall, a blue exercise ball, and a grey ottoman.</li><li>A nude woman and a nude man are having sex on a grey couch. The man is lying on his back, and the woman is positioned on top of him, facing away from the camera while looking back over her right shoulder. Her bare buttocks are centered in the frame, resting on his pelvis. A colorful tattoo is visible on her lower back and the top of her right buttock. The man's tanned chest and shoulders are visible in the foreground. The couch is adorned with grey pillows, and a grey shag rug is on the hardwood floor behind them.</li></ul><h2 id="how-was-this-made">How Was This Made?</h2>As with the original Anteros XXXL, we continue to believe there is too much secrecy around training. We are once again open sourcing the recipe with the hope that others will do the same.<h3 id="initial-finetune">Initial Finetune</h3>We used a custom training harness with the following hyperparameters:Batch Size: 120 LR: 2e-5 Precision: Original mixed bf16/fp32 Timestep MU: 0.906 Conditioning Dropout: 0.1 Optimizer: AdamW with zero weight decay L2 Regularization: OnThe L2 Regularization is the only real innovation here. Using the ideas from [Explicit Inductive Bias for Transfer Learning with Convolutional Networks](<a target="_blank" rel="ugc" href="https://arxiv.org/pdf/1802.01483">https://arxiv.org/pdf/1802.01483</a>) the weights decay towards a fixed starting point - in our case, the original krea2 model. We use this as a replacement for AdamW's decay and found it to be much less destructive.We trained this for 3 epochs of the full dataset, resulting in Anteros XXX Krea2 Unaligned.<h2 id="alignment">Alignment</h2>We then used a smaller set of one thousand hand-captioned hand-selected images. We ran this with the same settings for 150 epochs. After that, we generated ten thousand image pairs with saved latents. Using a custom trained classifier, we selected the preferred image (an audit of a significant sample of the results showed the classifier met my preferences about 95% of the time). We used the techniques described in [Diffusion Model Alignment Using Direct Preference Optimization](<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2311.12908">https://arxiv.org/abs/2311.12908</a>) to further refine the visual appeal of the generated images.The result is Anteros XXX Krea2 Raw<h3 id="turbo">Turbo</h3>Finally, we performed a 2 epoch run using the dataset's captions to perform the technique described in [Learning Few-Step Diffusion Models by Trajectory Distribution Matching](<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2503.06674">https://arxiv.org/abs/2503.06674</a>), which is cited by the Krea2 paper. We used the aligned model as teacher and the original Krea2 Turbo model as student.We once again used our custom training harness, this time with the hyperparameters below:Batch Size: 120 LR: 1e-5 Precision: Original mixed bf16/fp32 Timestep MU: 1.15 Optimizer: AdamW with zero weight decay L2 Regularization: On (against the student's original weights) Teacher CFG: 4.5 Steps: 8 "Fake" LoRA Scale: 10 "Fake" LoRA Rank: 256 "Fake" LoRA Alpha: 256<h3 id="hardware">Hardware</h3>All training was performed using four H200 GPUs on a single host using FSDP.

Anteros XXX Krea2

If you find you can't quite get the result you want from the Turbo model, and are okay spending more time on generation, give this a try. It's the result of our post-finetune alignment and should have more flexibility than Turbo. This is also a good starting point if you want to train your own character LoRAs.SetttingsSampler: Euler Scheduler: beta57 Steps: 40 CFG: 4.5

Files with SHA256: 8f0f358595db6993af73c5c918ef23c4328a1be30afa1e710d125b2b087f3987