Negative Embedding (Textual Inversion) for SD1.5 models
One technique for generating high-quality images is to effectively use negative prompts. However, rather than listing the same quality tags every time, Negative Embedding can be used to control the negative prompt very effectively. Of course, it is possible to generate high quality images without using Negative Embedding, but it is recommended to implement it to make prompt construction easier.
Please refer to the previous article for a detailed explanation of how to introduce the system, if you have not already read it.
Negative Embedding is available for SD1.5, SDXL, and Pony models. In this article, we will introduce Negative Embedding for SD1.5 models.
Negative Embedding (Textual Inversion) for SD1.5 models
Sample images are used to facilitate visual understanding of the effects of each model. These images were generated by applying only the negative prompts introduced to clarify the effects of the models and do not use worst quality, low quality,
etc.
EasyNegative
- Creator: 🔗gdgsfsfs
- Download: 24.7 kB
- Upload: 2023/2/1
- Trigger words:
EasyNegative
EasyNegative” by gdgsfsfs is a popular model at cibitai. It is trained using “Counterfeit”, so We think it is a good match with Counterfeit. In addition, we introduce demo images with “AbyssOrangeMix2_sfw” and “anything-v4.0-pruned”, so we think they are also compatible with it. The link is also available on cibitai, but we have chosen to link to the author’s original Hugging Face link.
There are two down load files, EasyNegative.pt
and EasyNegative.safetensors.
If there is no particular reason, download the EasyNegative.safetensors
one.
badhandv4
- Creator: 🔗Euge_
- Download: 18.92 kB
- Upload: 2023/3/8
- Trigger words:
badhandv4
Euge’s “badhandv4” is a model that supports the generation of hands, which AI typically struggles with. Some Negative Embedding models may alter the art style, so in such cases, it’s recommended to lower the model’s weight. It is advised to use a CFG scale of 11 or below.
Bad_NegAnatomy Textual Inversion
- Creator: 🔗EnthusiastAI
- Download: 210.85 kB
- Upload: 2023/5/7
- Trigger words:
BadNegAnatomyV1-neg
EnthusiastAI’s “Bad_NegAnatomy Textual Inversion” is a model that supports the generation of body joints.
bad_prompt Negative Embedding
- Creator: 🔗Nerfgun3
- Download: 24.92 kB
- Upload: 2023/5/2
- Trigger words:
bad_prompt_version2-neg
Nerfgun3’s “bad_prompt Negative Embedding” improves overall quality, similar to EasyNegative. Since the art style may become slightly anime-like, adjust the weight accordingly.
negative_hand Negative Embedding
- Creator: 🔗Nerfgun3
- Download: 24.93 kB
- Upload: 2023/5/2
- Trigger words:
negative_hand-neg
Nerfgun3’s “negative_hand Negative Embedding,” like the “bad_prompt Negative Embedding,” is a popular model on cibitai. It is an easy-to-use model with minimal impact on the art style.
Deep Negative V1.x
- Creator: 🔗lenML
- Download: 225.92 kB
- Upload: 2023/1/20
- Trigger words:
ng_deepnegative_v1_75t
lenML’s “Deep Negative V1.x” is a popular model on cibitai that improves overall quality. It brings out colors vividly, so you can achieve nice color tones even without using VAE.
Bad artist Negative embedding
- Creator: 🔗Nick
- Download: 7.08 kB
- Upload:
- Trigger words:
bad-artist
またはbad-artist-anime
Nick’s “Bad artist Negative embedding” is trained using “Anything-v3.” There are both normal and anime versions. Refer to the sample images above and download the style you prefer.
BadDream + UnrealisticDream
- Creator: 🔗Lykon
- Download: Bad Dream:207.85 kB, UnrealisticDream: 111.85 kB
- Upload: 2023/5/22
- Trigger words:
BadDream, UnrealisticDream
Lykon’s “BadDream + UnrealisticDream” are recommended to be used together. “BadDream” pairs well with “DreamShaper,” while “UnrealisticDream” is suitable for realistic styles. You can download “UnrealisticDream” by switching the model version.
veryBadImageNegative
- Creator: 🔗yunleme
- Download: 30.92 kB
- Upload: 2023/3/20
- Trigger words:
verybadimagenegative_v1.3
yunleme’s “veryBadImageNegative” is also a popular model on cibitai. It helps create well-balanced images overall.
Conclusion
This time, we introduced Negative Embeddings (Textual Inversions) for the SD1.5 model. While mastering negative embeddings won’t completely solve the AI’s struggles with hands and joints, it can often lead to satisfactory results. We encourage you all to utilize this technique.