Type | |
Stats | 2,634 0 |
Reviews | (169) |
Published | Jan 24, 2025 |
Base Model | |
Training | Steps: 7,000 Epochs: 27 |
Usage Tips | Strength: 0.8 |
Hash | AutoV2 1C21DFFD1A |
Trained on 200+ very highly detailed high resolution (min 2000px) images + 35 hires vidclips of amateur female face portraits, with visible skin pores, skin hairs and blemishes/moles.
The lora is intended to use for higher resolutions like 768px and up (width and/or height) - as training concentrated image and vidframes samples between 768 and 1152px this is where performs best.
My samples posted are made with DPM++2M, Scheduler Beta for 28 steps on the original tencent bf16 models at 512x768 resolution @73 frames, lora strength 0.85. The quants/fpX/fastvid I tried all produced visible different and degraded quality IMHO.
V1.1 works better with other Loras - put it around 3/4 weight of the other Lora; e.g. this Lora at 0.65, the other Lora (character etc.) at 0.9
For SkyReels or Hunyuan ImageToVideo vids, lower the strength to about 0.4 for vids over 73 frames. the Lora was only trained on snippets <73 frames, and this seems to have more visible (anti)effect on longer vids based on I2V model than T2V.