This lora is intended to fix bad geometry generated by Flux 2 Klein 9B when generating specific balloon shapes. It also reduces the likelihood of the model generating "foil" balloons, since these are much more prevalent in the dataset distribution than balloons. I was not happy with that, so I made this lora.
Conventional or smaller balloon shapes and pretty much all colours and transparencies are also possible, as well as unique combinations like gradients and a variety of inprints.
There was, unfortunately, some style transfer of the people holding the balloons if you go too generic, but loading one reference image can provide even better results than simply prompting the model.
Training:
A 60-image dataset was built using Nano Banana 2 due to a lack of good data online and the free access I get to it using Hugging Face. Heart-shaped balloons are rarely seen inflated as much as this lora makes them be. This model was trained using default settings in AI-Toolkit, coupled with a higher batch size of 4 to better preserve the balloon shapes, as training at BS 1 resulted in poor geometry.
How to use (edit and txt2img only modes):
Be as descriptive as possible. For better results and ease of use, you can ask an LLM with vision capability inside comfyui (for example, using Ollama nodes) to describe only the subject, then concatenate this description with a generic prompt:
1 - (generic prompt) - Using the reference images, create a photo taken outdoors in a chinese temple on a sunny, clear day, depicting a
2 - (LLM description of the subject) - asian male with long dark hair tied back with a gold hairpin wearing layered traditional robes in white, yellow, and gold with intricate woven embroidery and ornate belt buckle,
3 - (generic prompt describing action and everything else) - blowing air into a giant yellow heart-shaped balloon. The balloon is much larger than he is, has a distinct heart shape with two lobes that transition smoothly into an elongated bulbous neck that ends in a knot. The balloon has a white imprint of stars on its surface. The balloon has a completely smooth surface, with no wrinkles or seams.
4 - (optional, improve subject consistency) - Keep the same clothing and subject identity from the reference image. Use normal body proportions for a typical asian male
Use a "text concatenate node" to concatenate these sentences into a single one, then feed it to your Clip text encoder:

Description
Initial release, 3000-step BS4 run
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.

