v1.0Twin sister of my already existing LoRA, trained on same dataset as latest version. Training params is as follows: <pre><code> "network_dim": 20, 
 "network_alpha": 20, #treat dim and alpha as one param. Their individual sizes influence amount of details able to fit inside of lora and their ratio alpha/dim is specifing where your lora will be most active during generation steps (alpha/dim&lt;1 more details, alpha/dim&gt;1 more color blobs on earlier stages)
 "learning_rate": 0.0001,
 "max_train_epochs": 10,
 "resolution": 768,
 "repeats": 10,
 "caption_dropout": 0.1,
 "gpu_index": "0", #you won't be able to use gpu without correct Torch+CUDA pair
 "optimizer_type": "AdamW8bit",
 "lr_scheduler": "cosine_with_restarts",
 "lr_scheduler_num_cycles": 1,
 "lr_warmup_steps": 100,
 "train_batch_size": 1,
 "gradient_accumulation_steps": 1,
 "max_grad_norm": 1.0,
 "save_every_n_epochs": 1,
 "save_last_n_epochs": 4,
 "mixed_precision": "bf16",
 "gradient_checkpointing": true,
 "seed": 42,
 "noise_offset": 0.03,
 "multires_noise_discount": 0.3,
 "timestep_sampling": "sigmoid",
 "discrete_flow_shift": 1.0,</code></pre> Base Apparently, newly emmerged Anima model is a completly different architecture from sd or sdxl, making it mostly incompatible with many good loras. However, training loras for it is not as hard, as one may think. Standalone trainer app from <a rel="ugc" href="https://github.com/citronlegacy/citron-anima-lora-trainer-ui">citron</a> is able to make aroud 5000 steps in 3 hours and is using around 5gb of VRAM. It is not as smooth in installation process as one may wish for, but still managable. Especially if you know/can bother to learn how to correctly install PyTorch for your specific version of CUDA driver of GPU. I myself figured it out just from couple of searches in google, so nothing it is not a rocket science.

Shoulder deep fisting [Anima]

shoulder-deep_4(img2img_more).zip

shooulder-deep-fisting.safetensors

v1.0

v4.0 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Trained with only img2img dataset (52 images) and double as long (20 epochs). Works better with more elaborate prompts. Still have problems with glitchy anatomy. v3.0 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">This one is actually just merge of two prewious (v1.5 from civit trainer and v2 with low dimentions). <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Results are better, though it appears that process of merging two tensors together even without any modifications (pure math, mimiking what happens inside of lora loader) still hinders quality. I suppose such is the price for compacting everything into one file instead of two separate. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">If you want better results, use specified loras separately. v2.0 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">This ona was trained on my machine with ai-toolkit trainer. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Set dimention and alpha to 4, used same dataset and 500 steps. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Better holds general pose, but, expectedly, fails with details. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">General Main idea of this lora is to generate pictures of fisting, when arm is almost all the way inside of another person (preferrably - elbow not visible and shoulder close to hole). Still laking, but already able to generate good images. And only similar one I managed to fine on this site is <a target="_blank" rel="ugc" href="https://civarchive.com/models/1722532/lost-my-keys-deep-fisting?modelVersionId=1949329&amp;sync-account=green">this one </a>from user @<a target="_blank" rel="ugc" href="https://civarchive.com/user/HaydenHoosk">HaydenHoosk </a>but it is specifically tailored for generating furries (not my style, deffinetly). LoRA is trained with semi-syntetic dataset. Halve is made by passing real images from "orange youtube" through depth+openpose controlnets and other halve was generated with my supervision. I am still new in this field, so I don't quite understand wich settings of trainer and qualities of dataset impact what, but I will be improwing this model in the future.If anyone has good explanations or observations, I will be wery glad to hear them in the comments.

Shoulder deep fisting

Dataset only from repainted images and due to this is slightly smaller than previous ones. Works better with more tags.

Files with SHA256: 17b4e29656079addee0ad55f6b69705962a8bb118e6071a80a562e949c4bea04