moss_image_sft_4_steps_lora_v2_600_avg_rank_72_bf16.safetensors

更新:2026-10-19 20:49|发布时间:2026-10-19 20:49
0
0
0
moss_image_sft_4_steps_lora_v2_600_avg_rank_72_bf16.safetensors. Experimental MOSS-image related models SFT model can produce clear vertical banding artifacts, applying the later b…
头像
Gugugaga
类型
LoRA
基础模型
MOSS
发布时间
2026-10-19 20:49
文件签名
70796585d8d089279d1fd513679bfccf8d82f1716b4c96eef54688d436e9c860
moss_image_sft_4_steps_lora_v2_600_avg_rank_72_bf16.safetensors
311.93 MB

Experimental MOSS-image related models

SFT model can produce clear vertical banding artifacts, applying the later blocks with negative strength can be enough to fix that and reduce the overblown SFT look.

Full SFT: image

21-49 rank 512 SFT LoRA at -0.5: image

Or just use base with full SFT diff as LoRA, these ranks aren't enough to fully capture the SFT training.


moss_image_sft_4_steps_lora_v2_600_avg_rank_72_bf16.safetensors

Rank reduced (fro 0.99) and saved in bf16, source:

https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps/blob/main/euler-update-600/checkpoint-update-600/lora.safetensors

作品

暂无作品