Frequently Asked Questions
Q: What is One Shot?
A: One shot is an online service that will generate a character LoRA model for Stable Diffusion XL from publicly available video.
Q: What is a LoRA model?
A: LoRA stands for "Low-Rank Adaptation" and it is a method for adding specialized knowledge to a more generalized machine learning model.
Q: How does it work?
A: LoRA is a popular and lightweight training technique that significantly reduces the number of trainable parameters. It works by inserting a smaller number of new weights into the model and only these are trained. [ SOURCE]
Q: Can I see an example?
Q: What can I do with a LoRA character model?
A: You can use models from One Shot with any image generating software that supports SDXL and LoRA, such as ComfyUI.
Q: Is it anonymous?
A: Yes. We don't collect any personal information beyond what is required by law in order to properly manage the sale of tokens and we do not store any data that links token use to token sales.
Q: How would the payment appear on my bank statement?
A:
Q: Do you store my data?
A: Not for long. Your job history and downloadable files will stick around for a few days-- long enough for you to grab a copy and make sure you're happy with it. If you'd prefer a more traditional account, with login, password, and permanence, let us know!
Q: What is Video to LoRA?
A: The One Shot Video to LoRA service works by executing a number of AI-powered steps in a single automated pipeline. First the video is downloaded and individual frames are extracted. Then the humans in those frames are detected, classified, and sorted. The most prevalent human is selected and images of that human are extracted in both close-up and wide-angle shots. These images are then captioned in preparation for training. The images and captions are then trained against either the FLUX.1-D base model (for 2000 steps) or the Stable Diffusion XL 1.0 base model (for approximately 3000 steps across 3 "epochs") to generate the new model. Optionally, One Shot can generate four sample images based on the new LoRA for you to evaluate.
Q: What is LoRA to Image?
A: The One Shot LoRA to Image service allows you to evaluate a LoRA created by the Video to Model service. One shot uses a Large Language Model (LLM) to combine your prompt with the LoRA's tags to produce unique and detailed prompts for image generation using the selected SDXL "checkpoint" model.
Q: What is Synthetic Merge?
A: Synthetic Merge is an advanced feature offered by One Shot LoRA that allows you to combine multiple LoRA models into a single, customized model. Using our "synthmerge" process, you can select the LoRAs you want to merge and provide a trigger word or phrase. The system then generates 100 images based on the selected LoRAs, picks the most similar ones, and trains a new LoRA using these images and the original LoRA weights. The result is a brand-new LoRA model tailored to your specifications, along with four sample images. The entire process is streamlined and costs just 10 tokens.
Q: What are jobs?
A: Any time you ask One Shot to do something in exchange for tokens, One Shot will kick off a "job" to perform that task. These jobs will appear in your Job History and if anything goes wrong we will need your Job Number to track it and fix it for you.
Q: What is a Permalink?
A: The permalink allows you maintain a relationship with One Shot over time. Since we do not store account information or require you to "log in" there is no other way to "come back to" your tokens or your job history without the permalink.
Q: What are these tokens?
A: One Shot is powered by Newbound's TXCU tokens, which provide a layer of abstraction and anonymity between payment and service. TXCU stands for “Transmit eXchange Compute Unit.” A TXCU token represents the notional value of the underlying compute costs required to launch and run an online service. We wanted to call them “Compute Unit Notional Transfer” tokens, but marketing said no, legal advised against it, and HR gave us a stern talking-to.
Q: What happens if I don't save my permalink?
A: You will have no way to access your job history and you will lose any unspent tokens.
Q: I lost my permalink can you help?
A: No. We literally do not know who you are.
Q: I had over 2 models created, now I can’t find them!
A: That is not a question.
Q: It’s been over 4 hours and the Lora isn’t ready-- what should I do?
A: Contact us immediately, and include the job number.
Q: I paid 10 bucks and I didn’t get my tokens-- what should I do?
A: Contact us immediately, and include your name and the PayPal Transaction ID if possible.
Q: Can I gift my tokens to my friends?
A: Not at this time. Let us know if that is a feature you would like!
Q: Can I withdraw my tokens or exchange them for cash?
A: Not at this time. Let us know if that is a feature you would like!
Q: How long does it take to train a LoRA model?
A: Approximately 90 minutes. It depends largely on the source video’s resolution, the number of other humans (not the subject) in the video, and the amount of frames in the video where the subject is too small or not present at all.
Q: What happens if I use a link that isn’t a video or to a video that’s 100 hours long?
A: Your job will fail and you will have 5 fewer tokens.
Q: "No video detected" - What does this error mean?
A: The server attempted to download the content at the link provided, but no video was found. This can be caused by many things, even when there is clearly a video there. Many of our servers are managed by Amazon, and sometimes the IP address assigned to that server has been blocked by the video's host. Links to Tik Tok, for example, will often (but not always) fail.
Q: "No faces detected" - What does this error mean?
A: The video was successfully downloaded and processed, and humans were found in the video, but the faces on those humans were too small to train on. The cut-off is a minimum of at least 256 pixels in either height or width.
Q: Why do you caution against upscaled video?
A: The results are objectively bad. Go ahead and try, it's your two dollars.
Q: How much video is enough? Is more better?
A: Anything more than 8.3 minutes is overkill, unless it adds meaningful variety. WE WILL ONLY EXTRACT 500 FRAMES, NO MATTER HOW MUCH VIDEO YOU PROVIDE. We will not extract more than one frame per second, so anything less than 8.3 minutes will result in fewer than 500 frames being extracted. We recommend a minimum of 5 minutes worth of video, but good results can be had with as little as 1 minute.
Q: 500 frames are extracted, you say? The JSON file in the ZIP archive says, "Found 222 bodies and 96 faces." What gives?
A: We extract the maximum of 500 frames BEFORE we crop and sort them. Not every frame will have a usable image of your subject. For best results, you want about 300 bodies and 200 faces. Good results can be had with as little as 25 faces. Your mileage may vary.
Q: If I provide a video with two or more people in it, will the LoRA be some weird mash-up of different people?
A: No. We select the most common face and ignore the other people entirely.
Q: Does One Shot work on anime or cartoons?
A: No.