The new version is the final version. I made this a while back, I think it added some sex positions but I don't remember. It is trained on 2x the original data and should be better for realism.
This model attempts to achieve a higher level of realism than chroma can achieve on it's own. It was trained on 5 different types of images, but some trained better than others.
Facials: The model was trained on facials, out of the 5 things it was trained on it does these the best.
Deepthroating: The model does this relatively well, but can take some work to get right
Buttshot: Model does this well, not much stipulation
Handjob: Model often has trouble with this, can take a lot of work to get right
Squirting: Model either does this well or horrible, usually well but can have bad generations.
The prompting can be a bit difficult on this model, so it's best to download the sample images and use those prompts. The model works well with the chroma flash lora.
Description
Base version, unsure if I'll train further, only if people want it.
FAQ
Comments (26)
Legitimately interesting. Using this with goontune both at 0.35 create crisp and creative enough outputs I believe I can abandon flux realism loras. All I've been using for general style is "2014 photo of". My negative is pretty massive though.
Nice, thanx a lot. In your workflow integrated in the image I see that you are using a lora name "chroma-unlocked". What is the purpose of that one?
Probably just a speed distill: https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main
There are a set of distills with that name
@makiaeveli yup I used rank32 of that lora, it works without it
Please dont forget to publish metadata together with images. Not everyone uses Comfy to read your workflows.
the comfyUI nodes are the metadata. You click "copy" and you can search for "widgets_values"
,"widgets_values":["true realism, facial, photo of a young latina woman, face covered in cum, tongue out, winking at the camera, dark background, on a track, low light, casual selfie style, medium close-up shot, messy cum on her face, playful expression, natural makeup, slightly disheveled hair, candid moment, amateur photo"]}]
The JSON will also include the models and loras as well. Generally searching for ".safetensor" or ".gguf", or just "widget_values" and read them all
lol
@makiaeveli Do you think everyone who doesn't use Comfy needs to learn how to parse JSON in their head? There's an interface for that, and it's easy enough to just fill in a few fields if Comfy can't add metadata to images automatically.
@Unicom I assure you I used the search function when I copy and pasted the node data into this text box. I was simply providing a way to get the information yourself
What happened to the “waaahh don't criticize us it makes us cry!” comment thread? 🤣
real_new_final.safetensors?
It works great, I'm using it to add realism to some generations that lack it, but you either picked up the wrong file to share or trained it in a different model as it throws a bunch of errors when loading. At least in comfy.
It works fine for me, are you sure your loading it correctly? try loading other chroma loras
@FusionCow I am using other Loras, I trained quite a few as well that throw no errors.
I just updated Comfy and opened the chroma template and added a comfy core lora loader with to check and get about 30 errors similar to this one:
lora key not loaded: lora_transformer_distilled_guidance_layer_layers_0_linear_1.alpha
I also tried it using rgthree power lora lorader and load lora model only by it self and with other loras and it always throws this errors.
Either way and again, its a really good lora, thank you.
I hope you can reproduce the error on your end and figure it out as the errors are a bit distracting specially for someone who trains and tests LoRA's regularly.
@FusionCow I forgot to mention, I am using Chroma1-HD-Flash
@Mad_Pony_Interactive I've seen similar errors to this, it sounds like you're not using the right file, but that doesn't make sense. I honestly have no clue how you would even begin to reproduce this error, if you join lodestone rocks discord server there are a lot of people that know more than me
@FusionCow Yeah I am there. Ikr? It's really weird. To be honest I can't be bothered 😅 I'ts working. Was this with OneTrainer? I had similar errors picking up a backup instead of the model file or a save file, still that was throwing different errors.
@Mad_Pony_Interactive yes this was onetrainer, but I've never had any errors nor has anyone else i've talked to
@FusionCow Yeah, its strange, I only had errors with backup files as they are meant for internal use. Save files and model files are always ok. but don't stress, it's all good.
The LoRa is so great. What's wrong with it is that when you try to make it fall on the tongue, nothing ever remains, or inside the mouth.
This is a great LORA, it's bit tricky but with proper prompt crafting you can even generate some good stuffs.. I recreated nearly perfect screenshot of a BBC JAV with perfect vag penetrative action. I wonder what kind of dataset was used to train this.
I used a lot of images from nsfw forums and similar sites, i've actually had a newer version with more concepts trained for a while and just haven't released it, I might just release it tonight
@FusionCow where's it? XDDD
@ps1copato I completely forgot, doing it right now, I've been working on making a complete nsfw lora for wan, like over 500 hand captioned videos and thats been taking up my time
@FusionCow is it ready yet? Thanks again for your hard work! Keem the chroma stuff cumming.
@ctcf6 it's still training, getting iffy results, some concepts are learned well and some aren't
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.