V0.1:

Use strength 1-4. Can be used in negative to flatten+desaturate. If it's too saturated in the positive, use it in combination with this excellent saturation slider by baizuobusi to tone it down.

Fisheye + Depth for krea2. Dataset carefully selected as to train DEPTH and FISHEYE, photos that exaggerate the size of things close to the camera instead of just warping the image's perspective lines. Trained on REAL photography, no ai in the dataset.

Krea2 vanilla seems to miss the mark on the depth part of the fisheye effect(it looks more like a photoshop filter), so this lora is to push krea2 to learn that things closer to the camera should be bigger and things farther away should be much smaller, ie depth. Krea2 is a fast learner and this was pretty easy to create, the long part was finding all the best images. Of course, prompting correctly helps immensely. Put fisheyedepth in the prompt if you want the effect to be more pronounced.

This will be an ongoing series and is a WORK IN PROGRESS :), this is my first foray into 'utility' loras so any feedback is welcome. Look at images for examples/comparisons. Low file size and trained with low rank as not to interfere with loras. Version 0.1 adds both a fisheye effect and adds light+saturation, and was trained only on photography, but seem to work in any style, especially if you only let it run the first 2 or so steps--future versions hopefully will ONLY do the fisheye+depth effect without adding color.

It's not perfect, but the dream is to make a fisheye+depth lora that doesn't mess with YOUR (or my) loras.... but again, work in progress.

Forge users can set Bits in Low Diffusion to FP16(automatic) and run it only on the first x steps, for instance, <lora:v0.1-fisheyedepth_light_and_color:[[email protected],[email protected],[email protected],[email protected]]> will run it at strength 3 only on the first 2 steps (given you are doing an 8 step generation). This lets the lora stamp out the composition at the beginning then let your lora take the reins for the rest of the generation. Theres a comfy version of doing this but im not quite sure what it is.

Images trained + tested locally on forge neo. Trained on krea2 raw int8 model.

Krea2 is so good at listening that you almost dont need loras sometimes you just need the right prompt. Add this to your prompt with or without the lora to increase the effect even more:

"The image features an extremely wide field of view with pronounced barrel distortion, causing straight lines to bend into sweeping outward curves, especially near the edges of the frame. Objects very close to the viewer appear disproportionately enormous, with exaggerated proportions and dramatic foreshortening. Nearby surfaces expand toward the viewer while objects farther away appear progressively smaller, creating an extreme sense of depth and distance. The center of the image remains relatively natural, while the outer edges stretch, curve, and warp. Walls, floors, ceilings, and architectural lines visibly bow around the composition, creating an immersive, exaggerated perspective with dramatic foreground-to-background scale differences."