Doesn't require a trigger word, but adding them to the prompt will help sharpen the background and make it look like an actual iphone photo, otherwise it serves as an all purpose realism enhancer and will make all images more natural.
For Flux, the trigger is "iphone photo".
For SD3.5 Large it's "iph0ne", try lower weight around 0.5.. first attempt at sd3.5, it might be bad.
Trained with Ostris's ai-toolkit: Github page
SD3.5 Large V2 Test version comparison - FLUX V1 vs V2 comparisons below
FLUX
V1 is iPhone Photo 6-2000 /// V2 is iPhone Photo 7-15000
Description
FAQ
Comments (21)
i can use this lora v2 sd35 for gguf q4 turbo sd35 ? it will work ?
Hi, It would probably work with gguf version, but standard SD3.5 large, non turbo version.
The lora brings some blocks from Turbo into regular large model to make it faster. It would have unintended effects if used with turbo.
@Anibaaal why not creating the same but for sd35 turbo ? create it if- you can please .
I commend you for releasing this without early access because people would definitely give you buzz to download it. Great job.
Hi, How much vram is needed to train a lora for sd 3.5?
and what tool did you use?
Hi, during training it used around 18GB with option 'quantize: true' in ai-toolkit
SD3.5medium(Checkpoin) cannot use SD3.5Large(Lora) and SD3.5Large(Checkpoin) cannot use SD3.5medium(Lora),
Only SD3.5medium(Checkpoin)+SD3.5medium(Lora) and SD3.5Large(Checkpoin)+SD3.5Large(Lora) can be used.
First of all, "v2 SD3.5 TEST" Lora is not universal for SD3.5medium (Checkpoint).
"v2 SD3.5 TEST" Lora was trained with SD3.5Large. Then when I use SD3.5medium (Checkpoint) and use "v2 SD3.5 TEST" Lora (3.5Large), there will be a Lora weight problem and "ERROR lora" will be displayed. diffusion_model.joint_blocks.0.context_block.attn.proj.weight shape '[1536, 1536]' is invalid for input of size 5914624"
For those who have questions about this issue, I can upload a Youtube test video
Hi, yes, this lora is for SD3.5 Large only, it was trained before 3.5 Medium was launched, and the models have different structures so they aren't compatible sadly.
I have updated the post to clarify it's for the Large version.
Thanks for the heads up.
Great work, thanks!
Any plans to add a SD 3.5 Medium version?
My laptop struggles to run the large model.
Version mismatch on forge, doesn't works.
I'm also having issues. I absolutely cannot get this to trigger. All my other LORAs work fine.
Would you say V1 is more realistic and V2 leans more towards fantasy? Just based off the examples provided here. Also I think its weird that an iphone photo lora has these sci-fi examples, what planet are you from? :D Love the loras
Thanks :) V2 was trained slower and longer, maybe I changed the captions, I don't remember well but they use the same dataset images. The fantasy settings are because I wanted to show how the lora performs with diverse prompts, not only real world images, but yeah it is a little weird hahah
I think the idea is to make the sci-fi photos more realistic.
It's great to see the effect of the lora in fantasy images, for me it's better to understand the overall effect!
IMO this should be a recommended approach for lora creators to show before/after in different type of styles.
Also for me, v1 it' more amateur (juts by loooking at the examples).
Anyway! Thanks for posting your work!
hey, when i run iphone lora on a flux model (tried various) i instantly get an 'server error'. I use invokeAI, anybody knows how i can solve this issue?
same problem here
This Lora is just MAGIC! A must have
will this work with z image turbo?