wikeeyang/Ernie-Image-Flex-V1

Ernie-Image-Flex V1.0

本模型为百度 ERNIE-Image 的微调模型，相比 ERNIE-Image-Turbo 在以下几方面有所改善和提升 (The model is a fine-tuned version of Baidu's ERNIE-Image, with improvements and enhancements over ERNIE-Image-Turbo in the following aspects)：

1. 基本去除 ERNIE-Image-Turbo 8 步模型油腻较重、噪点过强、颜色偏黄的问题。(Essentially eliminates the issues of heavy oiliness, excessive noise, and yellowish color found in the 8-step ERNIE-Image-Turbo model.)
2. 相比 ERNIE-Image 需要 cfg=4.0 / steps=50 出图很慢，本模型只需 cfg=1.0 / steps=16 - 25，且可使用 Turbo LoRA 加速至 step=8，有更好的弹性尺度。(Compared to ERNIE-Image, which requires cfg=4.0 / steps=50 and is very slow to generate images, this model only needs cfg=1.0 / steps=16-25, and can be accelerated to step=8 using Turbo LoRA, offering better scalability.)
3. 更好的人像构图与人脸优化，画面更干净，复杂构图时肢体相比 ERNIE-Image-Turbo 不容易崩。(Improved portrait composition and face optimization, resulting in cleaner images, and limbs are less likely to deform in complex compositions compared to ERNIE-Image-Turbo.)
4. 中英文文字渲染能力得到较好延续，比官方原版略有损耗。(Continues to render Chinese and English text well, with slightly less accuracy than the official original version.)
5. 量化模型采用全新的分层/分块/混合精度量化机制，使 8bit 模型与 bf16 模型效果更接近。(The quantized model uses a new hierarchical / block / mixed-precision quantization mechanism, making the 8bit model closer in effect to the bf16 model.)

本模型依旧为纯净底模，可用于微调或 LoRA 训练，也可叠加各种 Turbo / Distill / NSFW / SFW 的 LoRA 使用。(Pure base model, suitable for fine-tuning or LoRA training, and can also be combined with various Turbo / Distill / SFW / NSFW LoRAs.)

感谢: @亮亮rayne , @GuangyuanSD 在本模型调制优化过程中给与技术、构图、美学等方面的指导和支持。

Much thanks to: @亮亮rayne , @GuangyuanSD for their guidance and support in technical, compositional, and aesthetic aspects during the finetuning of the model.

感谢：@T8star-Aix，@AI-KSK，@黑鹤001，@大麦茶看AI，@joeMultimedia , 山竹老师AIGC , CommintAI , 雪煮松风 , 冰缘排骨 , 活力网TV , HyLan_L , 等一众大佬对本人微调各种模型的技术评测、使用介绍、社区推荐及进一步微调提升等方面的建议和支持。

Much thanks to: @T8star-Aix, @AI-KSK, @黑鹤001, @大麦茶看AI, @joeMultimedia, 山竹老师AIGC, CommintAI, 雪煮松风, 冰缘排骨, 活力网TV, HyLan_L, and many other experts for their advice and support in technical evaluation, usage introduction, community recommendations, and further fine-tuning improvements of various models I worked on.

基础工作流请参见(Base Workflow)：v1-example-workflow.json

Also on: Modelscope

The model under Apache 2.0.