Upload 2527064/2840118/README.md with huggingface_hub

2527064/2840118/README.md

@@ -35,29 +35,61 @@ tags:
 
 ## Description
 
-**This LoRA can only do I2V of Frieren.**
+*(ALL previews are I2V. V2 previews made from lazy Chroma outputs and some others I quickly found on CivitAI and Booru images. Most previews are single-shot gens but some were selected as the best of 2~3 generated samples)*
 
-Jk.
+This LoRA is mostly meant to animate good sexual input images (I2V) with basic sex animations using very simple and short prompts and/or to assist other motion-specific loras.
 
-Multi-concept experimental LoRA. Work in progress. Or not?. Don't expect it to become better because I'm already hitting diminishing returns.
+This LoRA wants sex. It wants it so much that sometimes it gives you 'ok' animations even without a prompt. Its extremely biased in that regard so its really not good for advanced use cases such as changing characters positions or scenes.
 
-*(previews are not cherry picked and use super basic captions because part of the experiment is to see how well LTX-2 adapts to short captions - so yeah don't complain unless you can do better EDIT: kinda forgot to mention cause its extremely obvious but all previews were done in I2V mode - as recommended for this LoRA. Also it can do more than just hentai ofc)*
+If you give it a sexual scene input image that you can tell with just a glance 'what kind of movement the characters should be doing' - than this LoRA should be able to figure that movement on its own with just the correct trigger words.
 
-Trained without audio but had audio weights (from other loras) injected at low, balanced ratios - post training. Mostly for automated female moans though that pretty much doubled the size of the file. Whatever.
+**Audio**: not trained at all. It contains audio layers from other loras merged together then injected into this one.
 
-Separate T2V datasets (with WAN-like captions) and I2V datasets (with very short, movement-focused captions) were used with modified training code to ensure I2V mode exclusive and 100% for I2V datasets.
+V1: may be better at anime
 
-Meant for 2/3-stage I2V workflows. Body horror with T2V - most of the time. Input should be proportionally resized with width = 512 for 3-stage workflows and 1024 for 2-stage workflows.
+V2: better at realism, less over-fitting overall and broader knowledge
 
-Terminology: penis/pussy/fellatio/insert instead of dick/vagina/blowjob/penetrate
+**Terminology:** **penis/pussy/fellatio/insert****instead of****dick/cock/vagina/blowjob/penetrate**
 
-Main triggers:
+### Trigger words with notes (colored by how good the model is at it in a I2V context):
 
-- having sex in the missionary/cowgirl/reverse cowgirl/doggystyle position
+* having sex (this one is powerful due to its versatility)
+* having sex in the cowgirl/reverse cowgirl/congress/reverse congress/mating press/missionary/doggystyle position (if only 1 of these scenes exists in the input image than you should use this instead of just 'having sex')
+* having sex while lifted in the air (for scenes where the female's legs are in the air and should dangle around during sex)
+* performing fellatio (not very good. you should probably pair it with a fellatio-exclusive lora)
+* performing a handjob with her right/left hand
+* performing a handjob with his right/left hand on his own penis (male masturbation. V2 exclusive)
+* masturbating with a dildo
+* masturbating with the fingers of her right/left hand
+* thrusting in and out of her pussy (helpful to complement sex and masturbation scenes)
+* causing a stomach bulge with each thrust (only knows a little bit of this so if you want this concept you should pair it with an actual stomach bulge lora)
+* spreading her pussy with both hands // the fingers of her right/left hand (not good at this either but it does have some knowledge)
+* pubic hair // shaved pubic hair (this is useful if you pair it in T2V mode with other loras. you can add it to the negative prompt on workflows with CFG higher than 1 to significantly reduce the chance of it appearing)
+* horsedong (not much but some e621 data is in my datasets and I used this word instead of 'penis' in cases where it made sense. it can help in furry animations that contain this kind of genital)
+* inserts into her pussy (this was used for penis, fingers, dildos and other objects)
+* comes out of her pussy (did not use this for penis removal but it may help with that. it was mostly for fully/half inserted big objects coming out from a pussy - to teach the model pussy physics of such animations)
 
-- penis thrusting in and out (forgot to include this in the previews lol)
+### Extra trigger words (mostly for negative prompts in CFG>1 workflows):
 
-Many concepts haven't been captured yet - you can see in the previews that masturbation and handjobs are not working at all depending on the scene and sometimes it works even without prompting for it. Also it sucks at fellatio.
+* static image, no movement. (used on images during training)
+* the camera is slowly panning left and right // up and down // in circular motions
+
+### **Downsides/Weaknesses:**
+
+* Too much data to fit in a rank 128 LoRA, causing over-fitting. I believe what I tried to accomplish may be possible at rank 256 with batch\_size = 16 training but that's something I can't do.
+* In I2V mode this lora will ignore most of your prompt if its not something in the available trigger words. It highly depends on the input image and prompt but this can be heavily mitigated by not using distillation or reducing it and using CFG>1.
+* Camera panning bias. It likes to move the camera around a little even when not prompted for it. Can be mitigated with negative prompting - see 'Extra trigger words' section.
+* Breast movement bias so strong it may cause flat chests to move when they should not to the point it may even affect males depending on the input image and your prompt. May be solvable with negative prompting + CFG>1.
+* Cowgirl sex scenes with only the woman moving are very difficult to generate due to the lora's 'having sex' bias which in turn means 'man thrusting a penis inside a pussy' - as far as the lora understands it. This happens even without mentioning 'having sex' in the prompt.
+* Trained on very, very little anal sex data so it may hurt anal animations a bit. V1 probably does this more than V2.
+* Cant draw good genitals from scratch (T2V).
+
+### **Recommended workflow info:**
+
+* Option A (RECOMMENDED): 3-stage workflow with input image proportionally resized to a max dimension of 512.
+* Option B: 2-stage workflow with input image proportionally resized to a max dimension of 1024.
+* For input images with very complex backgrounds: use Option B, otherwise stick to Option A. Doing Option B helps avoiding unwanted motion of background objects and stuff because LTX sometimes does that if it does not fully understand the input. But this lora performs the best motion in 512px buckets so avoid Option B.
+* For better I2V results avoid distillation or set it to just 25% + CFG=1.5 + 12~16 1st stage steps (2nd stage steps should be 75% of stage1 and 3rd stage should be 50%)
 
 ---
 
@@ -67,4 +99,4 @@ Model: [CivitAI Model Page](https://civitai.com/models/2527064?modelVersionId=28
 
 Archive: [CivArchive Page](https://civarchive.com/models/2527064?modelVersionId=2840118)
 
-<!-- Version: 20260424_all_update -->
+<!-- Version: 20260515_upload -->