The latest version of the model was trained on the new Z-img base model and will work on both the Z Image Base & Turbo versions. BUT to get it to work on Turbo you will probably need to increase the strength to 2.0, at 1.0 it was having almost zero effect for some reason.
Made as a user request. If you don't know what this is, Urban Dictionary is here to help: http://torpedo-tits.urbanup.com/16189773
A LoRA trained to generate woman with Torpedo Tits / Conical Breasts. This model was trained on a curated dataset of 136 images with light/minimal captioning for 3000 steps. Generates a decent pair true to the concept most of the time and seems to work across different sizes and nipple types.
Main Trigger Word:
a woman torpedo shaped breastsAdditional Trained Concepts
puffy nipples
close up shot
front view
low angle shot
side profile
three-quarter view
See example images for prompt examples.
What it's capable of:
Multi-Style: Works well for photo style gens and surprisingly very well for Anime considering there was none in the training set but you may need to turn down the strength to .7-8 for anime generations.
Consistent Framing: Generates the concept well at most angles of framing
Archetype & Setting Variety: Can generate a wide range of female archetypes and integrate them into diverse settings/environments
Body Diversity: Responds very well to descriptors for age, body shape, and unique features.
Description
This version was trained on a curated dataset of 136 images with light/minimal captioning for 3000 steps.
FAQ
Comments (10)
Unfortunately this falls apart with other loras. Either you have to tune down the weight so much that it's effect is almost entirely gone or both loras together absolutely destroy z-image's quality.
That's pretty much the case with every single lora I've ever tried with z-image unless you use a block weights node. https://www.reddit.com/r/StableDiffusion/comments/1pthc20/block_edit_save_your_loras_in_comfyui_lora_loader/
But this lora will likely be disproportionately effected as is because it's trained to 3000. It's not overcooked on its own but it definitely would be combined with others. I can release one of the ~1200-2000 step versions of the Lora that may play better with multiple loras. I'll do that later tonight.
@TheAIDoctor cool thanks!
Even with this being the only Lora, it creams out the render, dilutes saturation, and actual torpedo shape not pleasant at 1 strength. Real bad with Euler/Simple which I think is most common combo so I tried the ecclectic mix in your samples and really not much better. Needs work.
I haven't experienced those issues in most of my testing but I'm going to release one of the earlier step versions to see if it plays nicer with multi-lora and more set ups. I use er_sde sgm_uniform for my generations, cfg 1, 10 steps, denoise 1.0
ComfyUI recently had an update supposedly correcting how LoRAs work with ZIT models, proving ComfyUI still has massive maths issues. Complicating this is how ComfyUI defaults to different maths paths depending on what generation of Nvida card one has. I myself saw one run give a completely different output from a second run with the exact same parameters in the same ComfyUI session (recent version/Blackwell). And Sage Attention 3, for those unfortunate enough to have installed it, is a disaster (use an earlier version).
Of course, many people claim ZIT is hopeless with two or more LoRAs, which may be down to how they are trained, or again point to faulty maths in ComfyUI. In my testing, this LoRA works fine on Blackwell with Sage Attention 2.
@blobby99 To further prove your point. I'm using Blackwell with Triton in my workflow and on an older version of Comfy. I've noticed on a couple of my lora's people mentioning that they can't get the same result from my prompts and seeds which probably is due to the differences in comfy and attentioning versions
@blobby99 I think it does really boil down to how the lora was trained, as I have my favorites and more often than not am running 4 loras at .4 and the fith lora being a light slider I usually have at -.4 which I understand the -.4 does not subtract but also adds so im usually at about 2 strength in total with minor tweaks here and there.
Zimg Base already! nice work, this lora is great
I try to work fast, but pretty easy since its just a retrain and all the dataset is already prepped.