<h1 id="description">Description</h1><p>Steers the model towards the style of gpt image 2. It improves details, lighting and texture a lot, <span style="color:rgb(134, 142, 150)"><em>though the image might end up looking slightly sloppy</em></span>.</p><p>This is meant as an enhancer so play with lora strength to allow more style diversity :)</p><p></p><hr /><h1 id="note-on-versions">Note on versions</h1><h2 id="moss"><span style="color:rgb(64, 192, 87)">MOSS</span></h2><h3 id="v0.2"><span style="color:rgb(250, 176, 5)">v0.2</span></h3><p>I found out better sampling settings than for v0.1, so comparing the images between the two wouldn't really be fair.<br />Still an overall improvement over v0.1, though just like Anima v1.1 it might be suffering from more saturation and more over detailing than the actual dataset. </p><p>Should work quite good as a quality / detail boost now.</p><h3 id="v0.1"><span style="color:rgb(250, 176, 5)">v0.1</span></h3><p>Trained on MOSS base for roughly 2000 steps with a high resolution portion up to 1344 at the end.</p><p>Not quite as good as I wanted, I most likely should have left it training a bit longer. You can push the strength up to 1.5 without too much side effects.</p><p>I will redo a run while shifting the timesteps a bit less so the model can focus on details more.</p><h2 id="anima"><span style="color:rgb(64, 192, 87)">Anima</span></h2><p><span style="color:rgb(250, 176, 5)"><strong>v1.1</strong></span> is closer to the images in the dataset, it adds a lot more texture and details, shifts the colors less, and affects the pose / composition a bit less too compared to v1. It has the downside of sticking more closely to the body and face of gpt-image-2 (which gets boring fairly quickly), and having some of the less pleasing repeating texture pattern of gpt2 image.</p><p>You might just prefer the effect that <span style="color:rgb(253, 126, 20)"><strong>v1</strong></span> give if you want less texture or punchier colors.</p><h3 id="misc-notes">Misc Notes</h3><p>This is trained without a reg dataset, so the trigger word is not needed, but helps ever so slightly.</p><p>I found some better optimizer settings, and fixed some bug in the custom code I added to the trainer I used between v1 and v1.1. That was most likely holding back all of my anima LoRA. Which explain the large change between the two version when trained with the exact same dataset.</p><p>I might train another version since the v1.1 LoRA was still learning when the save published was auto-generated, but it started to get bad composition in the following epochs. I might need to lower the shift during training to it focuses more on the details. Though it will never be "perfect" because Qwen VAE is one of the worse VAE alongside the old SD model ones.</p>

Style > gpt-image-2

gpt-image2_MOSS-base_v1.safetensors

MOSS v0.2

Files with SHA256: 637b98ee1d1a1bd878c7c3216b79a772abf181d2e6a313598956dcf35ae8dd4a