主要收录 gpt-image-2 图片生成提示词 和 seedance2.0 视频生成提示词。
Language models fundamentally calculate the probability of each sentence in natural language using a mathematical model. When you input a question to an AI, the AI calculates its answer based on probabilities. 2. Further...
The third stage is the Pre-trained Language Model (PLM), which is a natural language processing model trained using a large amount of text data. Relative to NLM, PLM utilizes unsupervised learning methods, eliminating th...
The fourth stage is the Large Language Model (LLM). You can understand the current LLM as a PLM trained on an exceptionally large dataset. For example, GPT-2 has only 1.5 billion parameters, while GPT-3 has a staggering ...
Compared to the previously mentioned SLM, NLM has stronger learning capabilities due to the power of deep neural networks, possessing better generalization and adaptability when learning the language model. For example, ...
This is the first stage of a language model, also known as a Statistical Language Model (SLM). Its basic idea is to establish a word prediction model based on the Markov assumption, which predicts the next word based on ...
The second stage is the Neural Language Model (NLM), which trains a model using neural networks to learn the associations and probability relationships between words. It can utilize large amounts of data for deep learnin...
Furthermore, modern language models are not limited to providing one answer to a question. In reality, for a given question, there can be multiple potential answers, and they are ranked based on their probabilities. Fina...
The simplest method is to use statistical approaches. In simple terms, probability is calculated based on the context provided as input. For example, in the sentence "Have you had dinner?", after "Have you had", there is...
Books: BookCorpus, which has been commonly used for smaller language models like GPT-2, includes over 11,000 e-books. It mainly comprises novels and biographies, with the most recent update dating back to December 2015. ...
"When I see you, my heart becomes colder below zero," meaning "half becomes cooler" (below zero sounds similar to "half" in Chinese).
Simply put, computing power refers to computational resources, or hardware. OpenAI has not disclosed the amount of computing resources used for training the GPT-3 language model. However, OpenAI's CEO hinted at hardware ...
"We need to go through Sunday," meaning Tuesday, Wednesday, Thursday, Friday, and Saturday (sounds similar to "we need to go through Sunday").
In practice, training large language models involves using a mixture of different data sources, rather than relying solely on a single corpus. For example, the data used for GPT-3 includes a combination of Books, CommonC...
From a data perspective, it is important to note that Chinese corpora have a small proportion in these models. ChatGPT is able to respond in Chinese because its training data contains some parallel Chinese-English data, ...
In fact, as we have seen from the previous explanation, one of the key aspects of developing large language models is data.
CommonCrawl: This is currently the largest open-source web crawling database. However, it contains a significant amount of dirty data, so the commonly used subsets are C4, CC-Stories, CC-News, and RealNews. Additionally,...
After understanding the principles of large language models, you may wonder how they are developed and what the key factors are. Recently, there have been numerous articles claiming that even 5G communication is crucial ...
Books: BookCorpus, which has been commonly used for smaller language models like GPT-2, includes over 11,000 e-books. It mainly comprises novels and biographies, with the most recent update dating back to December 2015. ...
If the content of the data is biased (for example, contains racist or sexist content), or even incorrect, the results generated by the large language model will also be wrong. 2. If there is little data in a certain doma...
Overall, the AeroGlide UltraSlim Smart Toothbrush by Boie is a highly advanced and effective toothbrush that is perfect for anyone looking to take their oral hygiene to the next level. With its advanced sonic technology ...
Of course, I'm only calling this a "temporary" situation, because in the future, robot + LLM combinations may be able to expand the range of computable problems a bit wider.
However, if this driverless car encounters a car blocking a one-way alley, it will be stuck, only able to call for help to get the driver to move the blocking car. This is a problem that cannot be solved mathematically.
For now, most LLMs can only solve mathematical problems. Some non-computable problems still cannot be solved, especially those requiring physical experiments.
Let me explain, the mathematical problems mentioned here are not just arithmetic operations, but problems that can be solved mathematically. For example, we can develop self-driving cars that can brake and stop by itself...
Overall, the AeroGlide UltraSlim Smart Toothbrush by Boie is a highly advanced and effective toothbrush that is perfect for anyone looking to take their oral hygiene to the next level. With its advanced sonic technology ...
One of the key features of the AeroGlide UltraSlim Smart Toothbrush is its advanced sonic technology, which uses high-frequency vibrations to break up plaque and bacteria on your teeth and gums. This technology is highly...
In addition to its sonic technology, the AeroGlide UltraSlim Smart Toothbrush also comes with a range of smart features that help you optimize your brushing routine. These include a built-in timer that ensures you brush ...
This can be literally translated as "hallucinations". Simply put, the model will speak nonsense. For some questions with little conversation data, after probability calculation, the model may spit out some very strange c...
For example, in linguistics domains, like translation or copywriting generation. Even if there are grammatical errors, there are a large number of similar usages in the data, so to some extent, the model's answers can be...
Simply put, this involves manually removing or reinforcing content from the data with human feedback. However, human intervention does not necessarily solve the problem. This method may lead to new problems, such as intr...
As I mentioned earlier, today's language models, even after four versions of development, still have not broken away from "probability calculation". Essentially, their core principle is still "using existing information ...
Of course, in the face of this limitation, what methods can be used to avoid it?
If the content of the data is biased (for example, contains racist or sexist content), or even incorrect, the results generated by the large language model will also be wrong.
The meaning of image prompts may be smaller, and I think the probability of being eliminated in the short term is greater. So I shelved the writing for the time being, and turned to study some auxiliary tools, which may ...
Sorry, you can search or google this by yourself,.
To be honest, I've written several, but the more I write, the more I feel that what I'm sharing is just some English words.
As of March 24, 2023, this tutorial has over 500 stars on Github. Thank you all for your support.
Starting this week, in addition to ensuring content updates, the site will also update a practice of mine every week. I hope everyone can explore various AI applications with me 😁
For typos, unclear parts, etc, feedback via github issue or our [Discord](https://discord.gg/B7Z7wjuUPg). 2. AI changes quickly. Outdated content can be flagged via github issue or our [Discord](https://discord.gg/B7Z7wj...
**Lower expectations: ** I'm no expert. I'm learning too, just a few steps ahead of you. There may be gaps or errors here. Also, for beginners, analogies or terms may be dumbed down a bit to aid understanding. Please be ...
🚧 : Contents need more work but don't affect reading. I'll improve later. - 🆘 : I need help here, e.g. prompt examples for a scenario. Ideas welcome via issues/Discord. - 🔴 : To prep you for advanced sections, basics may...
This is not a prompt collection. If you're expecting to find prompts to use directly, I'd recommend Googling instead. This focuses more on teaching methods and explaining why they work. - This is not meant as an authorit...
You are free to read, share, and modify this book as long as you keep the original author attribution and non-commercial license.
[AI 101](https://learningprompt.wiki/docs/ai-101): An AI tutorial for everyone. Still working hard on it. Stay tuned. 2. [ChatGPT Tutorial](https://learningprompt.wiki/docs/chatgpt-learning-path): From basics to advanced...
This tutorial has three parts. The content covers my journey of learning Prompt Engineering, summarizing some of the experiences and methods. If you are learning Prompt Engineering, I hope this tutorial can help.
[Midjourney Tutorial](https://learningprompt.wiki/docs/midjourney-learning-path): Through various examples, this teaches you how to use Midjourney. Even as a beginner, after reading this, you'll be able to quickly get st...
**Lower expectations: ** I'm no expert. I'm learning too, just a few steps ahead of you. There may be gaps or errors here. Also, for beginners, analogies or terms may be dumbed down a bit to aid understanding. Please be ...
[ChatGPT Tutorial](https://learningprompt.wiki/docs/chatgpt-learning-path): From basics to advanced, from a user's perspective, this teaches you how to use ChatGPT and how to use it even better.
Share — copy and redistribute the material in any medium or format - Adapt — remix, transform, and build upon the material
Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your...
No warranties are given. The license may not give you all of the permissions necessary for your intended use. For example, other rights such as publicity, privacy, or moral rights may limit how you use the material.
Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your...
Adapt — remix, transform, and build upon the material
[DiscoArt] If set, then `cuts-{step}.png` will be saved for each step, visualizing all cuts in a sprite image at each step.
[DiscoArt] The refresh rate of displaying the generated images in Notebook environment. The value has nothing to do with the rate of saving images and the speed of generation or sampling. It is purely about your browser ...
[DiscoArt] If set, then output will be saved as images. This includes intermediate, final results in the form of PNG and GIF. If set to False, then no images will be saved, everything will be saved in a Protobuf LZ4 form...
[DiscoArt] When specified, it overrides the default naming schema of the resulted DocumentArray. Useful when you have to know the result DocumentArray name in advance. The name also supports variable substitution via `{}...
[DiscoArt] Place text transformers of CLIP models on CPU. This saves more VRAM and will not hurt the speed at all on T4, P100, 3090; however, there are few community members report issue on V100 when it is `False`.
[DiscoArt] The group name of the cut-scheduling. For example, `default` corresponds to default: cut_overview: "[12]*400+[4]*600" cut_innercut: "[4]*400+[12]*600" cut_ic_pow: "[1]*1000" cut_icgray_p: "[0.2]*400+[0]*600" T...
[DiscoArt] Strategy when encounter misspelled token, can be 'raise', 'correct' and 'ignore'. If 'raise', then the misspelled token in the prompt will raise a ValueError. If 'correct', then the token will be replaced with...
[DiscoArt] A dictionary of string to boolean list that represents on/off of CLIP models at each step. CLIP Model schedules use a similar mechanism to cut_overview and `cut_innercut`. For example, `{"RN101::openai": "[Tru...
[DiscoArt] CLIP Model selectors provided by open-clip package. These various CLIP models are available for you to use during image generation. Models have different styles or ‘flavors,’ so look around. You can mix in mul...
The name of the batch, the batch id will be named as "discoart-[batch_name]-[uuid]". To avoid your artworks be overridden by other users, please use a unique name.
[DiscoArt] The number of samples generated at each steps. Say `batch_size=3`, then you can generate three images in one run. Not only this is faster than three runs, but it leverages loss function better and potentially ...
The schedule of inner cuts, which are smaller cropped images from the interior of the image, helpful in tuning fine details. The size of the inner cuts can be adjusted using the `cut_ic_pow` parameter.
In addition to the overall cut schedule, a portion of the cuts can be set to be grayscale instead of color. This may help with improved definition of shapes and edges, especially in the early diffusion steps where the im...
The schedule of overview cuts, which take a snapshot of the entire image and evaluate that against the prompt.
Sets the value of the clamp_grad limitation. Default is 0.05, providing for smoother, more muted coloration in images, but setting higher values (0.15-0.3) can provide interesting contrast and vibrancy. [DiscoArt] Can be...
eta (greek letter η) is a diffusion model variable that mixes in a random amount of scaled noise into each timestep. 0 is no noise, 1.0 is more noise. As with most DD parameters, you can go below zero for eta, but it may...
As I understand it, clamp_grad is an internal limiter that stops DD from producing extreme results. Try your images with and without clamp_grad. If the image changes drastically with clamp_grad turned off, it probably me...
sets type of Perlin noise: colored, gray, or a mix of both, giving you additional options for noise types. Experiment to see what these do in your projects.
Deep in the diffusion code, there is a random number ‘seed’ which is used as the basis for determining the initial state of the diffusion. By default, this is random, but you can also specify your own seed. This is usefu...
Normally, DD will use an image filled with random noise as a starting point for the diffusion curve. If perlin_init is selected, DD will instead use a Perlin noise model as an initial state. Perlin has very interesting c...
Option to use a secondary purpose-made diffusion model to clean up interim diffusion images for CLIP evaluation. If this option is turned off, DD will use the regular (large) diffusion model. Using the secondary model is...
Two alternate diffusion denoising algorithms. ddim has been around longer, and is more established and tested. plms is a newly added alternate method that promises good diffusion results in fewer steps, but has not been ...
Diffusion_model of choice. Note that you don't have to write the full name of the diffusion model, e.g. any prefix is enough. To use a listed all diffusion models, you can do: ```python from discoart import create create...
Saturation scale. Optional, set to zero to turn off. If used, sat_scale will help mitigate oversaturation. If your image is too saturated, increase sat_scale to reduce the saturation. [DiscoArt] Can be scheduled via synt...
Each iteration, the AI cuts the image into smaller pieces known as cuts, and compares each cut to the prompt to decide how to guide the next diffusion step. More cuts can generally lead to better images, since DD has mor...
Total variance denoising. Optional, set to zero to turn off. Controls ‘smoothness’ of final output. If used, tv_scale will try to smooth out your final image to reduce overall noise. If your image is too ‘crunchy’, incre...
Optional, set to zero to turn off. Used for adjustment of color contrast. Lower range_scale will increase contrast. Very low numbers create a reduced color palette, resulting in more vibrant or poster-like images. Higher...