REUPLOADED Correct v2 File On 6/17/2026: "Big_Breast_ID4_v2_fix"

v2.0

• Images are made in ComfyUI with a fairly simple workflow. I primarily prompt with natural language and have had good results. I think this lora is even better with the bbox prompting!

• No trigger word so be creative with language but medium, large, huge, and gigantic are the size categories trained. You can further influence size and shape with prompting

Changes:

• Dataset is still captioned with natural language (gonna wait to see if its worth redoing it all with json captions)

• Added more photos to dataset for more varied poses and size reinforcement (this includes smaller sizes)

• New training settings for more flexibility and diversity.

• Sizes are similar to previous.

• Better color and image detail

• Added: "lacating and spraying milk" for fun (it works some of the time)

Recommended Settings:

• Strength: Start at 0.5 to 0.65 for both conditional and unconditional. Anything near 1.0 is unstable but can give good results.

• Euler/Ideogram Scheduler

Feedback is always appreciated!

V1.0

Similar to my Flux 2 Klein LoKr.

• Ideogram 4 is very impressive. Tossed F2K dataset in and came out with decent results. Not well trained on pubic nudity but bare breasts and nipples work pretty well. The real difficulty was finding a checkpoint that could semi reliably get different sizes with basic prompting.

• Sample images are all made using natural language prompts. My limited json images looked great as well using the KJ node.

• No trigger word needed.

Recommended Settings (balance of clarity and size):

Dual Model (Conditional and Unconditional)

• Conditional strength (0.25-0.9)

• Unconditional strength (0.25-0.9)

I don't know the exact balance for the strengths. but given the LoRA is very likely overtrained i think a higher UNconditional strength "cleans" it up a bit...

IF, you are looking for bigger sizes then increasing the conditional strength to 0.9 should do the trick. The image gets fried around 1.0

• Use with the Realism Engine or Lenovo Ultrareal LoRAs etc. if you'd like to also have more flexibility with more themes. I typically run that at a low strength (Conditional 0.25-0.75) and (Unconditional 0.25-0.75)

• Dataset captions follow this general layout:

A medium shot of a girl with [size] breast(s) (implants) facing to the right. She is wearing [a tight, thin shirt] and [smiling]. Her breasts are [very perky] and [round]. The outline of her nipples peaks are visible.... [rest of prompt]

• Useful terms:

  • Size: small, medium, large, huge, gigantic, massive, etc.

  • Shape: round, perky, "bolted on"

  • General Modifiers: very, extremely, significant, prominent, etc.

  • Other: busty, bust, bustline, voluptuous, curvy, etc.

Just a v1 with basic training setting to get started. This is also a LoKr NOT a LoRA.

Please comment or post images. Feedback is much appreciated