This model is focused on improving the details in faces, skin, clothing, fur, and materials. It performs well in both txt2img and img2img/upscaling workflows.
Flux Model v2 - Weak vs. Strong:
The weak version is more flexible, allowing for a wide range of outputs.
The strong version enhances details more dramatically but has a higher risk of inconsistencies.
Both versions have their pros and cons. The strong model can really make details stand out, while the weak version offers more control and stability across different styles.
Recommended Settings:
CFG: 2.0 - 2.5
Steps: 20 - 30
SDXL/Pony:
Upscaling (1.5x - 2x) really brings out the fine details. I’ve personally had great results with full strength, but anything from 0.3 to 1 with 30 to 60 steps can make a noticeable difference.
Recommended Settings:
CFG: 2.0 - 7.0
Steps: 30 - 60
As a heads-up, most of the provided images (except for upscaled ones) were generated through txt2img without any editing, so there may be a few minor issues visible here and there.
Description
This version is now trained on 1000 high detail and high resolution images of Women, Men, Animals and Macrophotography altogether.
Best outputs are after upscaling, as the details really start to shine with 1.5-2x upscale.
i personally used full strength on the model, but everything from 0.5-1 and 30-60 steps makes a difference already.
Be aware, apart from upscaling, all provided images are only txt2img, so minor issues are sometimes visible.
For now this will be the last detailer update for SDXL and once SD3 will be updated and ready for loras, I will update this with the same dataset.
As always, id love to hear your feedback and ideas for improvement.
{
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 32,
"numRepeats": 1,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 0,
"noiseOffset": 0.03,
"targetSteps": 4000,
"enableBucket": true,
"networkAlpha": 32,
"optimizerType": "Prodigy",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": false,
"trainBatchSize": 5,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}FAQ
Comments (2)
I am having some very good results from this by only apply it in the last 0.25 or 0.2 steps. (Eg. Denise has been completed to ~80% of the steps then I reload the checkpoint with this Lora applied at 0.9 or 1 then finish the last ~20% of steps. 🧑🍳
In Auto1111 you can automate this with the Lora Block Weight or Dynamic Lora extensions
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.