SFT stripe artifact
Is this some kind of weird AI watermark? You can almost see it in the normal image with some of them.
Maybe its some kind of artifact that was trained in on accident from a synthetic dataset they used in training?
I'm having some luck countering this but it's not perfect: https://huggingface.co/MATLOWAI/MOSS-Image-SFT-DeBand/blob/main/images/presets_fog_forest_color.jpg
I think this might be caused by using the wrong vae. When i use the standard flux2 vae i get stripes in 30% of images but with the vae from this repo i havent gotten any stripes so far. Im still not quite sure about this tho so id appreciate if someone else could try it out.
This is probably because MOSS-Image distilled from gpt-image-2/2.5 — those have vertical colored-stripe watermarks/artifacts that look exactly like this. Frankly, a text-to-image model doesn't need that much gpt-image-2 distillation; I'd only consider it for an image-editing model. Hope the team can ship a post-training checkpoint trained on clean data soon to wash these gpt-image-2 artifacts away.
I've started another discussion with my fix. If people want to test it out! Thanks
moss_sft_destripe_top64_bf16.safetensors
I think this might be caused by using the wrong vae. When i use the standard flux2 vae i get stripes in 30% of images but with the vae from this repo i havent gotten any stripes so far. Im still not quite sure about this tho so id appreciate if someone else could try it out.
You mightve missed the first example I posted showing the vae from this repo being used
Happens regardless of vae
Doesn't happen with raw model
Only with SFT model