AITK-CineFlux

Cineflux Model: Version 2

Hey everyone! I’m excited to share Cineflux V2, a big update with some really awesome improvements. This version uses the same dataset as V1.5, but pushed to 86 epochs, meaning you’ll get way finer details and much better realism. It’s all real screengrabs—no synthetic data—so the results are top-notch when it comes to cinematic realism.

One of the major upgrades here is the captioning. It’s much better at understanding cinematic concepts (though it’s not perfect yet—still working on that!). The only downside? It might lack some creativity from time to time, but I’m planning to fix that in the next iteration.

How to Use Cineflux V2:

If you’re using Comfy, I’d suggest trying my workflow, which is already set up with the model’s basics. But if you want to play around with settings, here’s what I’m using right now:

• CFG: 7

• Flux Guidance: 3.5

• Steps: 30

• Sampler: Euler Simple

Do’s & Don’ts:

Do:

• Be specific! The more details, the better. This model loves cinematic prompts.

• Use keywords like "cinematic," "film still," "photography," "film grain."

• You can even prompt for specific directors or movies—this model recognizes them! Here’s a list of directors and movies you can try:

  • Directors: Denis Villeneuve, Christopher Nolan, Alfonso Cuarón, Frank Miller, Gareth Edwards, Guillermo del Toro, Wong Kar-wai, Sofia Coppola, Ridley Scott, Sam Esmail, George Miller, Michael Bay.

  • Movies: Blade Runner, The Creator, Mad Max, and more! (Feel free to explore movies related to the directors above!)

• Try different strengths. I usually go for 1, but if you want a more subtle touch, drop it to 0.8 or even 0.4. You’ve got plenty of flexibility here!

Don’t:

• Avoid vague prompts like "cat chasing a chicken"—this model won’t respond well to that.

• Stay away from terms like "realistic," "photorealistic," or "8K UHD." These can actually degrade the output. Remember, cinematic doesn’t mean hyper-realistic!

• Watch out for hands! Sometimes this Lora struggles with them, but you can add negative prompts like "bad hands" or "poorly drawn hands" to fix it.

Prompt Template:

Here’s a basic template that works really well with Cineflux V2. It’s not fixed, but it helps give the model a clear direction:

• [Medium], [Character], [Clothes], [Environment], [Conditions], [Color Palette], [Lighting], [Camera Type], [Film Stock Type], [Director/Artist/Movies], [Additional Ambiance/Mood Keywords].

Example: “Cinematic, medium shot of a man and woman standing in a narrow, dimly lit alley, the soft glow of neon signs in red and green casting long shadows on the wet pavement, rain lightly falling, circa 1960s Hong Kong, shot from a low angle, soft green + deep red palette, shot by Arri Alexa using Kodak Portra 400, by Wong Kar-wai, the scene exuding a deep sense of longing and quiet romance.”

This template allows you to control everything from the medium (photography, cinema, illustration) to lighting, camera type, and even director/artist. You can be as detailed as you want with color palettes and mood, but you don’t always have to include everything—leave some room for the model to interpret based on your needs.

Final Tips:

Cineflux V2 works best with 16:9 or 21:9 formats. You can experiment with vertical shots, but the results might vary. And as always, feel free to share your results—I’d love to see what you create!

Cineflux Model: Version 1.5

Introducing Cineflux V1.5, an updated AI model designed to bring a cinematic flair to your creations. This version uses a new, hand-crafted dataset that enhances the model’s ability to capture specific directors' and artists' styles. Whether you're looking to recreate a particular cinematic look or experiment with artistic influences, this model is built to deliver.

Trained on a dataset of 1k images in the first version, this update refines those results with more precise captioning and control. While the best weights are still between 0.5 and 0.8 for optimal performance, you can now expect improved results thanks to the new dataset.

I encourage you to test the model, push its limits, and share your outcomes! Post your creations and rate the model—your feedback will help shape future versions. Let’s see how far we can push cinematic AI together!

Cineflux Model: Version 1

With the recent hype on flux i had to update my loras, this one was done on Civitai with 1k images, the captionning isn't quite what i wanted so the results are a bit rough.
during my testing phase i wasn't able to determine the best weight, 0.5 - 0.8 seems to be a safe bet though.

I'll update this model with a training on my GPU, it will become better overtime. I have over 25k real screengrab for the dataset, this should provide better understanding in currently lacking subjects.
Have fun with the lora, can't wait to see what you guys wil create with this.