Moss Alpha reduced dataset to about 23k images cherrypicked but trained in multi-res 512-768-1024, barebones but works decently nice with tags or NL not fully generalized to all contents in the checkpoint (finetuning Moss is hella expensive) used a tags = 30%, nl = 30%, tags_nl = 20%, nl_tags = 20% so should work with most types of prompting, use the same text encoder and vae as Moss-Base or Moss-SFT this checkpoint is based off Moss-Base, trained with SDNQ + INT8 with torch_varlen to ease the training of multi-res.

This is meant to show how Moss-Image learns, and it learns damn fast and damn easy

v2.0 reuses the same captioning and tags regime as v1.5 but with reduced data to 240k images due to constraints at 1280px but is multi res at 1024 and 1280px training somehow with more stable style across seeds and prompts, again all images have Natural Language captions, those Cross-Attention wasn't trained so https://github.com/An1X3R/Anima-Artist-Mixer will work as expected but at 1280px you hit a very big wall in speed and inference times, v2.0 preserves the ability to use latent styles without much dilution, model seems to like ModelSamplingFlux so samples were generated using it, mostly.

v1.5 a bit less stiff and less overwhelming style much less ai-generated style, all 400k images were trained with Natural Language and Tags on a rate of tags = 50%, nl = 10%, tags_nl = 20%, nl_tags = 20% so all images have Natural Language captions to preserve Anima's functionality on text and pure NL prompting, masterpiece/best quality are recommended still in pure NL.

Custom Mix of finetunes over Anima Preview 3 - Base, providing a familiar AI-generated Aesthetic into most seeds basically stabilizing to an extent the style variance over seeds, finetuned purely on 1024x1024 AR Buckets, the model aims to have its own default style without causing much dillution on latent characters or styles, issues from the base model have not been solved by this like multiple views being in negatives and it still happening on some wide shots, this model specifically trained some layers not all of them.

Usage Tips:

• Positive Prompt: masterpiece, best quality

• Negative Prompt: bworst quality, low quality, blurry, jpeg artifacts, bad hands, bad anatomy, signature, watermark, censor, bar censor, mosaic censorship, mosaic censor, blur, multiple views or similars.

Technical Settings:

• Sampler: ER_SDE, SA Solver Pece

• Steps: 20-28 / 12-18

• Scheduler: Beta, Linear Quadratic, Simple

• CFG Scale: 4.5-6.0

• Shift: 5.0 - 8.0

• Resolution: 1024x1024 or any aspect ratio that falls within ~~1.048.576 pixels,

If you want to support me, consider liking the model, and/or tipping buzz/donating to my Ko-fi, as this was finetuned on 8xH100s