'Improved Amateur Realism' - [STYLE] [LORA] [FLUX] - spectrum_0001 by 'AI_Characters'版本v7.0 (ID: 1293808)

v8: Major rework and appropriate name change of the entire model. Probably my last update to this LoRa for a while. I don't think I can do better for now and I want to move on to other LoRa's.

Works best with mundane, real-life scenes.

Recommended inference settings (v8):

Rule-of-thumb:

• lower guidance, lower amount of steps = more realism

• higher guidance, higher amount of steps = better coherence

This is however not universally true, as depending on seed and prompt this can vary.

• 1.0 LoRa strength (only lower as a last resort or when needed, e.g. when combining with other LoRa's)

• min. 2.5 (more realism) - max. 3.5 guidance (more coherence)

• 20/25/40/50 steps (these seem to be the optimal convergence points)

• euler + ddim_uniform (more realism) OR simple (more coherence)

Recommended upscale workflow (v8):

I highly recommend pairing this LoRa with a latent-upscale workflow, as the LoRa was trained with unusually highly detailed and high-resolution images, which the standard FLUX 1 megapixel resolution doesn't do enough justice of. You can find such a workflow by me here:

https://www.dropbox.com/scl/fi/g95wqgl1iktyupvl4rklu/two-times-latent-upscale.json?rlkey=en07fqp9qnicnadxlu2uzlskk&st=5j3x3ce6&dl=1

LoRa description:

This LoRa model was designed to generate images that look more "real" and less "artificially AI-generated by FLUX". It achieves this by making the natural and artificial lighting look more real and making the bokeh less strong. It also adds details to every part of the image, including skin, eyes, hair, and foliage. It also aims to reduce somewhat the common "FLUX chin and skin" and other such issues.

Training Workflow:

TBD

Current config: https://www.dropbox.com/scl/fi/1wxltd4e22ig16pkdspnq/True-Real-Photography-v5-STYLE-LORA-FLUX-spectrum_0001-by-AI_Characters-_20250110-012256.json?rlkey=moec02lx5bx7cvnvyvdjzsg7j&st=9l6q7z9d&dl=1

Then resized to 32 dim and merged to fp8 weights.

Disclaimer:

I do not claim to be the current owner, original creator of, person or legal representative of a person depicted in, any of the concepts that my models, including this one, aim to emulate nor the training data used to train this model. I do not claim that any of my models have received an endorsement by those respective individuals. I also do not claim that the emulation attempted by my models represents a 100% accurate depiction of the original concept in question nor do I claim that the quality of my emulation reaches the same quality of the original concept. All credit goes to the respective current owners, original creators, or people and I encourage you to support them in any way you can.

If you are the current owner, original creator of, person or legal representative of a person depicted in, any of the concepts that any of my models aims to emulate and would like that specific model to be removed, then please leave me a private message here or on Reddit with proof of authenticity of your identity and claim and I will remove it.
Additionally, I do not endorse my models being used to violate the law or enact still legal, but immoral acts such as the spread of misinformation using deep fakes.

Donations

All my public work will forever be free to use. Likes, downloads, and good reviews are already support enough.

However, if you would still like to support me financially, please do not donate Buzz to me, as I have no use for it other than promoting my work on the front page of CivitAI. If you have money to spare and want to support me, it is better to donate to my Ko-Fi: https://ko-fi.com/aicharacters

All money donated there will go towards covering my ~5000€ costs accrued over an entire year of me renting multiple RTX 4090 24gb and A100 80gb GPU's almost every day to train over ~1000 variations of test models with different underlying base models, datasets, and settings for SDXL, FLUX.1 [dev] and SD 3.5 Large in order to find the most optimal combination of base model, dataset, and settings.

Commissions

TBD

描述:

• KNOWN ISSUE: A few prompts, such as "A man in his late 20s is laying back on a grassy hill, wearing a hoodie and jeans, looking up at the sky with a thoughtful expression. The background shows a clear blue sky with a few clouds." return wrongly orientated images. I thought it might be related to one of the new images having wrong EXIF data orientation so I thought I fixed that and retrained, but the issue is still there. So no fix there yet, although it seems to affect only a very small number of prompts as far as I can tell. I will of course work urgently on fixing that!

• Based on a lot of feedback from a Reddit thread about my model (that I made), I have decided to completely revise it again, as v6 was clearly vastly inferior to v5. Luckily, this has resulted in the best model yet. By far.

• The model now has much more natural and detailed looking skin, less often or less strong occurrences of FLUX chin (at lower guidance values), looks much more real now, and overall is just better in every way. I would say that after 7 versions now and months of work, this is finally THE state that I always wanted to achieve and it can easily compete with the rest of the FLUX Realism LoRa's out there, if not even beat them imho

• What I did was

  • adjust dataset again: switch out the previous AI generated images with some previously unused photos of mine

  • change the trigger to "early 2010s snapshot photo captured with a phone and uploaded to facebook, " as extensive testing showed that this is the best one (I could come up with during multiple hours of testing)

  • remove "artstyle" from the trigger, as that caused some images to turn cartoonish or otherwise non-photoreal

  • switch to lower FLUX guidance values as the default for samples, instead of 3.5 more around 2.5, unless the prompt demands it

训练词语: early 2010s snapshot photo captured with a phone and uploaded to facebook,