Z-Image Turbo
Trained with over 650 images and 30,000 steps, this can produce decent results, but also monsters. Workflows are embedded for what I used for the samples. I used it with my Speedo Tan Line LoRA and set them to about 0.6 each otherwise I got trash. I used sa_solver/beta, 12 steps
seed_v3/beta is my current choice but that changes like every week
Trigger word: pen15
TLDR:
These ZIT LoRA were my first attempt at anything other than character LoRAs, so I don't have detailed tests as to exactly what worked, but here's the rundown:
I have started to do captioning through the QwenVL node in ComfyUI using the NSFW model, Qwen3-VL-4B-Thinking-abliterated (NSFW), using a workflow that batch processes a folder of images to resize and caption them
Many of these images for this one were captioned with Python code I modified to batch process image files using joy-caption-alpha-two. It did not do a good job at differentiating between flaccid and erect. I have another more focused dataset that I resized and cropped all the images to 512 squares and has all erect penises (coming soon), but this one surprisingly gave equal or better results for ZIT at least.
For my latest character lora, I followed the advice from this article with excellent results, and gave QwenVL these instructions:
Caption EVERYTHING you see except for the man and his hair or body type. That means outfit, backgrounds, lighting, camera angles, skin details, water droplets, etc. must all go into the caption. Refer to him as Brock
I used AI Toolkit and its adapter to train this on my 5090:
ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v2.safetensors
Description
FAQ
Comments (27)
can I ask how it was trained? de-turbo or with the adapter?
only flaccid? only cut? how did you caption it?
I've also been trying but my results are not as good as yours.
I just modified the writeup here to include that info.
(The LTX one "kinda" works for POV stuff, just FYI.) 👀
Like I said. Alpha version. I'm finding detailed prompts will help. Use Grok and tell it your problem and the prompt for LTX-2. It helped me big time
Even if it just keeps the penis attached to the body in I2V then it will be a major help :)
Like I said. Alpha version. I'm finding detailed prompts will help. Use Grok and tell it your problem and the prompt and it has helped me.
@coachbate I'm early into testing it but I generated a classic dick falling off LTX I2V vid then repeating the exact same prompt but just added your lora in and... SUCCESS. The cock stayed on him~! So thank you for this.
It doesn't. Floats away.
Great work! I'm getting great results in T2V, even much better than the examples you posted. I can get ultra-realistic and detailed penises. I'm using the prompt: bwc, penis, huge veins, detailed glans. Please make a (PUSSY / INNIE PUSSY) version. Thank you!
Works pretty well for I2V in combination with using image batches extracted from wan generations as the first initial frames. (Or gif frames, mp4 frames, whatever.) Your lora helps a ton with keeping the body and focus coherent even if the body part disappears and reappears. Thanks!
FWIW image batches guarantee motion in ltx2 i2v, and effectively biases the gen toward the motion depicted in the source frames on the fly without a lora. Can select singles like 0,1,2,3,4,5,6,7,8 or doubles like 0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7 to get 8n+1 source frames to inspire the model. VideoHelperSuite has a Select Images node and Load Video node, or use your favorite. Match the source framerate and target framerate to avoid slowmo/fastforward unless that's desired.
Please share some below. Knowing that it helps people will fuel future motivation for me.
can you post a mp4 that has the workflow embedded that shows what you mean? I don't quite follow but sounds interesting
@coachbate I'll try to post something after work sometime. You can drag an animated webp or gif into Load Image and it'll automatically treat it like a batch of frames. You can't feed arbitrary frame counts as input to ltx2 or you'll get an error, it has to be in multiples of 8, plus one extra. VideoHelperSuite has nodes that help with frame batches, I'm sure other packs do too. The idea is ltx2 will observe the motion and content in the first few frames and effectively extend the video from there, plus it adds audio. imo it improves the chances of a successful gen but it's still a diceroll with the model's terrible prompt adherence re: stuff it isn't trained for.
girls has penis too
that should probably work fine, just it isn't of interest to me, but feel free to add examples to the showcase
What he wanted to say was, if there is a girl in the image as well, she also will get a penis besides the man. So this is bad.
@SD_Agamemnon So my use of neutral pronouns gave everyone a penis? I have to use straight LoRAs all the time and lowering strength and prompting can usually work, or set a last frame image in your workflow. If there's a way to train it to work with gay, straight, trans, let me know and I'll implement it, but mine will cater to the gays since there are so few LoRAs out there.
Enables proper touching on I2V which would otherwise be impossible. Example below
Does affect character consistency, had to play around with settings/strength to keep the original image intact.
But overall great first Lora!
wonderful, thank you. There's never enough of this.
I always wanted a lora that could do various penis forms, like both big and small, cut uncut, hairy and hairless. I wonder if that's possible.
I also always wondered, what if you trained both a penis lora and bulge lora together, especially with video.
Could you get more realistic bulges? Like would having the penis data make it able to create more detailed penis outlines underneath underwear etc.?
@ForeverNecessary737716 I can try that because I've got several other concepts I'm training and one I called underboner, which is just erection bulges. I'm willing to try different things.
@coachbate ya, that's exactly what I'm so curious about as well, and I wonder if mixing nudes with under-boner would get more accurate shapes.
I'm actually trying youpenis lora right now and it's working very well, with clothes it's not 'under' but at least it spills out in a realistic way. Compared to other loras there are relatively few floating or inverted dicks.
@ForeverNecessary737716 I think it's so hard to get just the one concept working that two seems like it would confuse it. My underboner one was looking good with my test images and prompts but right before releasing I tested on more start images, and only 10% were doing what I expected. They all grabbed and smooshed up and massaged whatever was in the middle of the frame, like their partner's stomach which made him look like he was fat shaming and rubbing around his belly fat he pinched off. It's been moved to the back burner.
there are 2 penis loras out of Klein9b so far. What do you think if you combined datasets with those 2 and made the ultimate penis lora?
if they use the same model, you can load them at different strengths and blend your own. That would involve time and money for server compute time, and the authors all wanting to share their datasets. I know I personally do not want to share my dataset. I'm happy to share how I created it and some of my LoRAs.