Converts photos to a more RAW photo with neutral colours and lighting.
Trigger phrase is "Change to RAW photo".
Qwen Image 2.1
Use euler_a if the output is noisy.
Can also be used for T2I at lower strength; 0.3 is a good starting point.
Flux 2 Klein 9b [I2I / edit mode]
Increase CFG to increase the strength of the effect. I used CFG 1.0 for all the samples.
Flux Kontext
Increase strength according to how processed the input photo is.
v2
Improved conversion of colour graded photos
Improved handling of crushed blacks and blown whites
Note: I used Flux Guidance = 1.0 in all my examples.
Do share your experience and feedback!
Description
Trained with v2 dataset. Trained on base but seems to work well on the distilled version too.
FAQ
Comments (6)
is the idea that you convert to raw and then you can apply your own colour grading?
Yes
@olivetty So, make it look worse to make it look better?
@playtime_ai_ that would imply "raw" is a look, when it's not. Raw is just data. I'm not really sure what they think they're trying to achieve here because "raw" has nothing to do with this.
The starting inference image could have started life as a raw image as well and you would never know. Because it's not a look. Like, what's next, make your image look "LOG"?
how to solve the pixel shifting?any refer resolution to deal with the input image?
From the official Flux.2 code, I can see that there is a constraint that your image dimensions must be divisible by 16, which is likely the source of the pixel shifting. Try cropping both sides to a number divisible by 16.
Also, the maximum size supported seems to be 2 MP.