8/31/26 UPDATE: On my last trip, I dropped my iPod approximately 6 inches, causing the screen to spiderweb with a huge crack along the whole screen. Due to this, I will likely retire this lora. I have some stuff planned for a V3.0 once I retrain V2.0 on all the models, but that's likely it. The good news is, I found an iPhone 4s at a junk market and am working on gaining access to it, from there I will likely make another phone camera Lora based on the 4S camera, which is noticeably higher quality but still not great. should be fun!This model attempts to make your outputs more like what a 4th generation iPod Touch camera will output. This means Poor dynamic range, washed out colors and noisy shadows.I think this particular Lora differentiates itself because I took all the photos myself. I found an old iPod (4th generation) at a garage sale and took a bunch of pictures over the past month to create the dataset. This means the actual camera qualities should be baked into the model VERY well. the downside is that the dataset is pretty small (127 images currently) and limited to stuff a can actually take pictures of. This means it's pretty bad at nsfw.It also tends to give images a more Candid, Amateur Look. That's because the dataset is completely amateur and candid.MOSS ImageMoss his been a PITA to train for. shout out to @ObsidianDreams for sharing an AI-Toolkit config that actually worked.Strength 1.0-1.5 is good. up the strength from 1.0 if the effects come out too subtle (its somewhat inconsistent.Krea2This one also has a strong push towards a more candid, amateur style, especially as you go to higher strengths. I had to train this one on the turbo model since i don't quite have enough VRAM for the raw model.Subtle: 0.3 - 0.5 (safe, keeps composition similar to no Lora but adds some old digital camera smoothness, compression, and color washing)Dramatic: 0.7-1.0 (changes to a more candid and amateur composition, and outputs that look very true to a shitty old phone camera)⚠️ V2.0-SPECIFIC SETTINGSV2.0: Use a strength between 1.0–1.5. This version seems to benefit from a slightly higher strength because it was trained on Krea2 raw.This version does great with random objects, animals, and landscapes, but it will make your character a bit less photogenic at higher strengths compared to the base model. Note: This has been fixed in V2. characters look hot again.Another "strength" of this one is that it will push the style to realism. this can be useful in a lot of cases. the image of the teddy bear in the rain on the seed and prompt i used is actually more of a cartoon image on base Krea2, while with my Lora it looks a lot more realisticFlux2_Klein_9BThis version came out pretty good, much better than the 4B version. keeps composition super well! I also discovered you can prompt away alot of the effects this Lora applies to an image if you use too many "positive" tags (lighting, professional photoshoot, masterpiece, that kind of thing)subtle: 0.5 - 1.00high influence: 1.5 - 2.0I like to keep this one at around 1.5, but 2.0 is good too. Gotta figure out how to train FLux better so the Lora strength doesn't need to go so high. if anyone has advice, let me know!Flux2_Klein_4BProbably the most subtle version. you keep the aesthetics of the non-Lora version fairly intact with this one. Honestly i will probably retrain this one to be more dramatically iPod Touch 4th gen at a later pointFor subtle influence, keep the strength at 1.0-1.4=5For dramatic results, strength 1.6-3.0Z-Image BaseThis version is slightly more subtle. I would keep strength 0.8-1.0It also works for Z-image Turbo, but I would up the strength to 2.0-3.0Z-Image TurboThis version might honestly be a bit overbaked on the candid and amateur compositionI would keep strength between 0.3-0.8You will get your best results with plain Z-Image Turbo. I got some pretty terrible outputs on some of the Finetunes on here. it can work if you keep the strength pretty low.If there is actually some interest in this model, I can keep taking more pics in my travels and try to get a better dataset. I could also add nsfw pictures to my dataset if i can find them and confirm they were taken on an iPod touch 4th generation.I can also include the dataset if interested, but I will probably censor out any images of myself that I used in the dataset so I don't dox myself.

iPod Touch 4th generation Camera Style

I havent quite cracked MOSS Lora training yet, but this is by far my best version so far. gonna keep experimenting and see if I can get the size down and get it to the same level as the krea2 v2.o version

iPodTouch_4thgen_MOSS_v6_ObsDR_ver_000004000.safetensors

MOSS - v1.0

Krea2-v2.0

Files with SHA256: 724347043acf1220c51d975f9f4d17c50f309f4edb4333792f52c11915f7e841