This is tests of training a flux model, and all licensing conditions for the Flux Dev model apply. You can find the details here: https://huggingface.co/black-forest-labs/FLUX.1-dev. It's a great model!
A huge thank you to Datacrunch.io for generously providing their resources for training and support. I highly recommend them.
This is Checkpoint. CLIP and VAE is already on
This is my model trained on Flux, and it aims to create more diverse and realistic faces. The results show improvements in anatomy, facial features, and overall realism. However, a side effect is that noise appears in the images. This can be corrected with any upscaler - just find one that works for you. The model was trained on approximately 7,000 images over the course of several hours using powerful H100 cards. If the test results are mostly positive, I will train it on a larger dataset.
The number of training steps was limited due to budget constraints. Additionally, I used a small portion of my dataset, only 7,000 images, to reduce training time. When I get the opportunity, I plan to train a more refined model. You can use this model as you like, but please remember to adhere to the Black Forest Lab license terms.
I welcome any feedback in the comments - feel free to share any errors or shortcomings of the model. This is just an experiment.
Description
I would like to express my sincere gratitude to the teams at datacrunch.io and GoogleCloud for their invaluable support in developing the Flux_realistic_SaMay_v2 model. Thanks to their technology and the power of H100 GPUs, we were able to train the model on 3,500 images and reach new heights in ultra-realistic image generation.
Advantages of the Flux_realistic_SaMay_v2 Model:
Enhanced Realism: The model excels at handling complex lighting, textures, and shadows, making the generated scenes truly immersive and lifelike.
Improved Detail: Provides high-quality details in textures like skin, fabrics, and reflective surfaces.
Depth of Field: Offers excellent clarity in the foreground while maintaining realistic depth and atmospheric quality in the background.
Natural Lighting and Color Representation: Mimics natural lighting, including "golden hour" effects and realistic shadows, with vibrant and dynamic color representation.
Versatility: Performs effectively across various scenarios—from urban and nature landscapes to portrait photography—making it adaptable for different industries, including fashion and nature photography.
Vivid Contrast and Clarity: Creates high-contrast images where foreground elements stand out sharply against the background.
The Flux_realistic_SaMay_v2 model represents a significant advancement in generating photorealistic images with detailed textures, natural lighting, and vivid contrast. Whether you need urban landscapes, portraits, or action scenes, this model delivers high-quality, photorealistic images that can be easily customized for various creative projects.
FAQ
Comments (21)
Please add the GGUF version of your V2 model
I think OP is planning on adding it gradually, to get more visibility
@Polygon Hopefully
Yeah, soon. I need to test it
Please add the U-NET version of your V2 model
I think OP is planning on adding it gradually, to get more visibility
@Polygon you dont say!!
The preview images look really nice! Does this model refine only faces and people? Was any training/refining done to improve the details and textures of things like nature, interiors, cityscapes etc?
this model works mainly for people, of course there is nature and architecture in the dataset, but only as a background for people. I have trained a separate model for nature and architecture, but I need more time to test and possibly change it
@Sa_May Awesome. Thanks! I'm looking forward to trying out the nature/architecture model.
Please release Lora version too!
saw your post on reddit and was eager to try you rmodel, thanks for making it available.
I'd like to hear your opinion on the model
After getting some men out of the model who looked a bit...special, I tried "woman laying on grass" and found myself back in a world of SD3 monstrosities. Some of them were so weird that when I tried to post them, CivitAI flagged them as X-rated even though nothing a human would recognize as naughty bits were visible. I love the image quality of the output, so I hope the creators will work on anatomy.
Update: It seems that non-laying and non-horizontal images do much better. So as long as you stick to portrait orientation and sitting/standing it's great!
I will take your comments in the next version, post your prompts and I will check them too
@Sa_May Thanks for the response! I posted the images which include the prompts, which I kept very basic "woman laying on grass. film photo taken on SLR camera"
@epzer Thanks, I see your images, I didn't emphasize the model to the people kowtowing, but definitely this problem is not solved. With multiple generations you can get it right. You can train a small Lore on people lying on the grass - that will solve your problem
@Sa_May Ha ha, I'm admitted not that interested in people laying on grass, but I found that laying on other, more typical things like couches and beds also created issues. But I also want to say that this model has amazing image quality and, in many cases it seems, better prompt adherence than base Flux-Dev. Not sure how you managed that but I'm kinda happily flabbergasted by it!
The one place it might need some more data is in the wardrobe. It does not seem to reproduce fabrics well, or details in say the a design pattern on the fabric (artistic) when prompting for something like "...with a colorful floral design", even when passing through some loras. Otherwise, it produces more natural results. Very nice. Hoping you continue to develop.
I previously offered a mixed, but constructive comment, so I wanted to post another purely positive comment—which is that this model sometimes seems to have better prompt adherence than base Flux, which is wild!
You're so right. It's fabulous
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.