C00M - Cum [Flux]

This might be the one I am most satisfied with the result from so far!

Trained with 84 manually selected and cropped photos, all 1024*1024. The captions are simple and minimalist, and was done half-assed but are accurate. I introduced one single token as the trigger word, and used it in the middle of the captions:

• «Photo of a woman with C00M on her » followed by short and simple contextual clues like «face and in her mouth.».

This seems to have made the model very aware of what C00M is, and allowed it to generalise over to completely separate settings than the training data.

Not one single training image features a full face. Most are around 30% of a face, but there are a couple with a bit more, maybe up to 45%. This seems to make it play along okay with character LoRA’s, but I haven’t tested it out that much yet.

As usual with Flux, lower guidance coupled with more tokens yields better realism, while higher guidance allows for better prompt adherence with the tradeoff of polished AI-look.

Because I am a poor soul without a GPU, I play around with diffusion models on HF from my phone for the most part. Since my phone saves the outputs as .jpg without the metadata embedded, gallery photos are without this info unfortunately. Only exception is the output from the training run here on Civitai (the square one).

If you want to know, here are generation info:

• Camenduru/Flux.1-dev-diffusers on HF as base model. Don’t know if this differs from regular base dev or is just a mirror.

• Sampler/scheduler i dont know

• 28 steps

• Guidance 2-3.5

• LoRA str 1.15

Prompts vary a bit. What I have included sometimes is «opalescent stringy blobs of C00M» or «a mix of chunky blobs and thick viscous liquid C00M», but just C00M works well on its own.

It does NOT seem to want to output a mouth actually filled, only on tongue and dripping. Let me know if you manage it.

Also, please post your results below!

Training info (please let me know what I could improve/tweak):

• 84 pics

• 8400 steps

• 0.00010 LR

• 2 batch size

• Both networks 32

• Other options untouched