When you select a move, Pikachu releases thunder. When you switch, Squirtle comes out and attacks with water.
In this Pokemon battle demo, the game side calculates the result of the chosen move, and the AI turns that content into a battle animation. The game you operate and the video showing the results are connected. The use cases for video generation have expanded significantly from just making a single short work. Overseas public examples include movies where the audience votes on the sequel, classrooms that take questions, and hosts who pitch fictional products. On the other hand, there are works you can try from public prompts, such as creating filmed footage with different poses from a character image. This time, I will introduce 15 cases focusing on the creations and the published methods. Since individual works, official service demonstrations, and app development examples are mixed, I have added those differences in each section. All posted tweets have over 10,000 views as of September 10, 2026. The number of views is the number of times the post was displayed, not the number of video plays or viewers. The numbers are unified to the values of the same day, and generation time and costs are stated as reported by the authors. Public instructions should be read as materials for how to make them, and are treated separately from proof that the results met all conditions.
MiniMax H3 Series: Video Generation Models You Can Try from a Browser
MiniMax H3 is a model that creates videos based on text or images. H3 Max is a version further trained by fal based on H3, and H3 Max Turbo is a derivative that emphasizes speed and processing volume. Regular H3 also appears in this article, and the final long-form piece uses H3 Max Director. Director is a separate mechanism for continuous long-form generation and is not the same operation as regular short-form generation. You can try the first one from fal's browser screen. Choose Text-to-Video if creating from text, or Image-to-Video if using a starting image. After logging in, select a model and enter instructions or usable images. Before generating, check the duration, aspect ratio, resolution, and price display. If it's a public instruction that switches actions within 15 seconds, it's easier to compare if you read the duration as well. The conditions for the free trial and the paid generation screen are different, so prioritize the display at the time of execution. In the following games and programs, apps that handle voting, memory, and playback are also used. It's easier to follow the examples if you separate the part drawn by the video model from the part where the app accepts operations. In fal's official guide, you can try H3 Max 5 times a day for up to 15 seconds in the Free Sandbox. Since conditions differ from paid generation, check the remaining free count and price display before execution. I tried generating for free below:
0:00 / 0:12
According to the official display on September 10, 2026, it is 75% off until September 14. The price for 768p is $0.01 per second for Turbo and $0.02 per second for Max. Generating 15 seconds at 768p with Turbo would be $0.15. After the discount ends, it will be $0.04 per second, so it would be $0.60 under the same conditions. The end time and time zone for the discount are not specified. Please check the latest official guide when using.
1. Pokemon Battles Become Anime When You Choose a Move
2. A Generative Video Game Driven by Your Choices
3. 15-Second Video Generation Demo at MiniMax Design
https://x.com/AmControo/status/2094402543157702964. Individual introduction/generation demo. Confirmed Sept 10: 403k views / 318 likes. @AmControo's post is a generation demo using H3 Max at MiniMax Design for a 15-second video. According to the author, the time taken for generation was 33 seconds. The 15 seconds here is the length of the completed video, and 33 seconds is the generation time reported by the author. The two numbers represent different things. This can be seen alongside the next post by Alice as an example of outputting from a service's usage screen. Since the details of the video have not been confirmed here, descriptions of the subject or development are not added.
4. H3 Max Demo Showing Everything from Operation Screen to Output
https://x.com/AliceInfoAi/status/2094393221292200316. Individual introduction/generation demo. Confirmed Sept 10: 267k views / 257 likes. Alice The Ai Expert introduces the process of using MiniMax Design along with the generated video. There is a display indicating a wait for generation in the middle, followed by a video using a Western-style hat, cards, fire, etc. Unlike posts that only show the finished video, you can follow the flow from the usage screen to the result. If example 3 is an example of reading the author's report of 15 seconds and 33 seconds, in example 4, you can focus on what was output. This is a demo showing Western-style themes along with the actual usage screen. The screen recording of the post is about 30 seconds, but the introduced generated video is 15 seconds. The length of the screen recording, including waiting time and operations, is not counted as the duration generated by the model at once. Even when using fal as an entry point, the usage limits and screen displays of MiniMax Design are not necessarily the same. Even for the same Max introduction, you can check which service it is being used on.
5. Island Survival Show Where the Audience Moves the Story
https://x.com/renoiseai/status/2095353704480395534. Service official program demo/trailer. Confirmed Sept 10: 205k views / 146 likes. People standing on a beach, scenes of gathering wood, a bonfire, and SOS drawn in the sand. Finally, it connects to a video overlooking the island. Renoise released a trailer for an AI survival show where the audience operates the story. The official service explains that they are using H3 Max Turbo via fal. In the short video, you can tell the setting is a deserted island and the people appearing are spending time there. Beach life and collaborative work are shown as the show's themes. The approximately 17-second trailer does not show the voting screen or the operations before and after the choices are reflected. Audience participation is the author's project explanation, and I have not confirmed the current operating status or the action where voting results are reflected. The relationship is that the app accepting input passes the audience's participation to the story, and Turbo creates the video. It is not an explanation that Turbo alone handles voting or show progression. Compared to the game in example 2 where a player chooses a single action, this is a project where the audience participates in a show. Detailed instructions for production are not public, but it is listed as an example of expanding video into an audience-participation program.
6. A New Video is Created Every Time You Scroll
https://x.com/Rames_Jusso/status/2094160939776155706 https://x.com/fal/status/2093844097148559588. Individual app development example. Confirmed Sept 10: 160k views / 642 likes. Short videos flow on a vertical smartphone screen. As you scroll, it changes from a woman in front of a mirror to a cloth bag, a dog in a kitchen, and a scene opening curtains. The appearance is a feed where short daily videos appear one after another. James Russo released it as a TikTok-like app that generates content as you scroll. It looks like a TikTok screen, but what was introduced is a self-made app. According to the author, the video being viewed can also be extended up to 30 seconds. In the approximately 21-second public demo, you can see the process of sending screens where short videos are lined up. H3 Max is used for video generation. The author explains that OpenAI's Codex was used for the work of creating the app and demo screen records. The production time of a few hours was also reported by the author. Creating short vertical videos with fal is the video generation operation. The part that prepares the next video and the part that creates a screen you can send with your finger is separate app development. What can be understood from the original post is the finished image of the screen and the separate use of H3 Max and Codex. In examples 1 and 2, the operation of choosing a move or action was connected to the next video. Here, the operation of 'sending to the next,' which is the same as normal video viewing, is the trigger for generation. This is an example showing that entry points for video generation can be created other than screens where you enter instructions every time. The author's explanation that it can be extended to 30 seconds is a function to continue the one being watched. It can be read that the operation to send to a new video and the operation to see the continuation of the current video exist separately.
7. Turning a Character Image into a 10-Pose Fashion Video
8. A Classroom Where an Explainer Anime Starts on the Spot When You Ask a Question
9. 'Pixelshop': AI Host Introduces Products When You Hand Them Over
https://x.com/pveerina/status/2095257520634040574. Individual app development example. Confirmed Sept 10: 112k views / 419 likes. Prasanth Veerina's 'Pixelshop' is a shopping show-type app where products are accepted and an AI host introduces them in a stream. The author explains a mechanism that accepts products, introduces them in the program, and also responds to chat questions. This is an example of connecting video of talking about products with viewer input. The posted screen shows a host in blue-green clothes and a background with TVs. In the scene handling an electric toothbrush, the product display and the chat column on the right are also visible. Unlike videos that only show the host largely, the product, program, and chat are on the same screen. It's a form where you can follow what is being introduced from the screen without relying only on the host's words. The video model used is H3 Max Turbo. Product acceptance, chat, and stream progression are functions on the app side combined with video generation. In example 11, viewers participate in the program by thinking of non-existent products. Pixelshop can be read as a different example in that it highlights the mechanism for introducing accepted products. From the public post, the finished screen connecting product introduction and viewer questions can be understood. Since the reproduction procedure for the entire app has not been confirmed, I will not infer and add the surrounding technical configuration.
10. 'THIS WAY': The Movie's Sequel is Decided by Audience Vote
https://x.com/henrydaubrez/status/2094865361250533465. Individual participatory movie/app development example. Confirmed Sept 10: 97k views / 1,255 likes. Dark forest, flashlight, video following a person. Next to it, two actions for the audience to choose are displayed. Henry Daubrez's 'THIS WAY' is a prototype of a movie where the audience chooses the next development by voting. In the original video, the screen for the movie viewers and the screen for handling progression are lined up. Whether to stay with the character Mara or follow another voice. Whether to choose the left path or the right. Choices that force a judgment in the middle of the story appear on the screen. If it were just forest video, it would be a horror short, but the feature of this work is that it accepts audience choices next to it. According to the author's explanation, it has a mechanism to remember characters, the flow of the story, and results from past choices. Furthermore, it generates two futures in advance and adopts the one chosen by vote as the story that follows. The role of creating video and the role of deciding which story to proceed with are combined. In the clock tower adventure in example 2, one player chooses an action. In THIS WAY, it is presented as multiple viewers deciding the next of the same story. Therefore, not only the video generation speed, but also when to put out choices, when to close them, and how to pass the results to the next are part of the program. From the public post, you can even confirm these production thoughts. What you can take away from this work is the method of making where there is another unchosen future behind the movie visible to the audience. Preparation that cannot be understood from just the played piece becomes visible through the author's explanation. The process of maintaining character names and past events is also app-side design. There is preparation for how to pass the story so far to the video to be generated. The Max version of the post and the subsequent update to Turbo are treated separately. Features added later are not made to have been completed at the time of the posted original video.
11. AI Host Sells Fictional Products
12. How to Use Turbo from a 15-Second Generation Example
https://x.com/umesh_ai/status/2095365608938545392 https://x.com/fal/status/2095210540083884453. Individual example/public prompt. Confirmed Sept 10: 28k views / 286 likes. Rounded letters and bright-colored 3D animation. A and apple, B and ball, C and cat, D and duck appear in order, and finally the four letters line up. Umesh's post emphasizes generation speed and cost, but if you read the prompt along with the video, you can also understand the use of time. Public prompt guided by the creator. In the instructions, a star effect is put first, followed by A, B, C, D, and a final review. The time to use for each letter, the shape and color of objects, and gentle camera movements are written. You can follow this order in the generation results as well. Videos displaying letters are themes where it's easy to confirm not only if the atmosphere matches, but also if they are the intended letters and order. Note that the linked prompt was released on August 3, and the author guided it from the Turbo example on September 3. This video is a new Turbo generation result using previously public input. The author reported generating 15 seconds at 480p in less than 4 seconds, and the cost was $0.09375. If using this video for education, check if A-D are correct and if the audio reading matches in the finished product. Since you can check the time-order specification against the actual display, it is also suitable as a theme to start reading public prompts. If you look at the first half, middle, and end of the original text in order, the introduction for each letter and the scene showing them together at the end are separate. It is more concrete up to how to proceed with the 15 seconds than instructions that just place A-D on the screen. The point of re-guiding prompts released in the past from new Turbo examples is also helpful. The release date of the prompt and the model that created this output can be confirmed separately.
13. Applying Koda's Shooting Prompt to Turbo
14. Comparing Sci-Fi Combat with the Same Character and Instructions
https://x.com/Mayz1169/status/2094759305895694774. Individual comparison example/public prompt. Confirmed Sept 10: 21k views / 147 likes. White armor, red eyes, mechanical legs spreading from the back. Kiki released a comparison of SF combat between H3 Max and Seedance 2.5 using the same character and instructions. The video is H3 Max on the left and Seedance 2.5 on the right. On the left side, you can see a close-up of red eyes, movement using a wall, and a scene landing and holding a weapon. Even with the same settings, the way the person is depicted differs. The H3 Max side has prominent flat anime lines, and the Seedance 2.5 side has strong shading on the face and armor. In the reply, character settings and video instructions are separate. This is an example where you can read what to move and how to move separately. On the character side, white/black/red color scheme, female pilot, 4 mechanical legs, 2 missile pods, etc., are specified. The head armor also has a sharp shape reminiscent of cat ears. The video side is a 15-second battle set in an industrial facility. It is specified to show mechanical legs and weapons in the beginning, followed by an attack, movement using a wall, landing, shooting, and a final pose. It even describes how much time to use for each action, such as 0-2.5 seconds for the initial deployment of mechanical legs and 13.5-14.5 seconds for the final shooting. In the video, mechanical legs, red eyes, red attack trajectories, a figure by a wall, and a scene holding a weapon can be confirmed. Comparing the instruction's timetable and the video, you can follow the scene showing equipment and the scene fighting separately. The shooting instructions in example 7 were to change poses and angles while maintaining the person's appearance. In example 14, the number and shape of equipment are further specified, and it proceeds to actions using that equipment. When preparing a reference image, you can also check if not only the face but also the equipment you want to move is shown. This is to see if you are trying to move weapons or mechanical legs not in the image while only leaving the original instructions. Which of the left or right you prefer also changes depending on the art style you want to create. In this article, I introduce this as an example of comparing how characters are drawn and how scenes proceed, without making the other model's side a result of Max. Public instructions for character settings Public prompt for video
15. From One Prompt to a 13-Minute Continuous Anime
https://x.com/OdinLovis/status/2095601237375009146. Individual long-form example/public prompt. Confirmed Sept 10: 16k views / 131 likes. Lovis Odin introduced that they created a 13-minute anime in one continuous generation. According to the author, the production cost was $78. The model used was H3 Max Director. This must be read separately from the short-form generation by regular Max or Turbo so far. The public instructions describe the protagonist's appearance, clothes, and art style. A man with a small mustache and a light-colored jacket with coral-colored trim are specified. The art style includes a 3D expression with watercolor, ink lines, and a grainy feel like printed matter. The content is to move through worlds with water, plants, and animals while keeping the same protagonist. Walking, running, flying, swimming. Scenes move according to the protagonist's actions. The confirmed extracted images also have a man in a cream and coral jacket, scenes by the water or garden, and scenes with animals. It is a configuration that repeatedly uses the costume and person in a long video. I have not concluded that the person was perfectly maintained throughout the 13 minutes. What I confirmed here are the public instructions and scenes extracted from the video. Also, one generation does not mean a one-shot without cuts. The public prompt also has fast cut switching and screen splitting effects. Continuing one generation while changing scenes and shooting methods within the video. You can see that the number of generation tasks and the number of cuts in the work are separate. One more point, the original text has a specification of 30 scenes of 15 seconds. The total is 7 minutes and 30 seconds, so the 13-minute output introduced by the author and the time specification do not match exactly. Rather than taking it as 'it will definitely be 13 minutes if you use the public prompt,' this is an example of reading how they try to carry over the character and art style for a long time. While example 14 allocated combat actions to 15 seconds, example 15 is an instruction to maintain the person and greatly change the environment. Even for the same H3 series introduction, the length and configuration they are trying to make are different. If you read the long original text, you can follow in the order of protagonist specification, art style, and scene sequence. Instead of reprinting the full text, I will leave a link to the input released by the author. Public prompt for Director H3 Max Director Official Guide
Start with One that Has Public Input
If you want to try moving a character image, you can read Koda's instructions in example 7 and Jin.B's applied example in example 13 as a set. If you want to make characters with equipment fight, it's the SF combat in example 14. After seeing the original post, open the public prompt and check the model name, input image, and duration. Prepare an image you can use and check the price display before generating. If it's an example using an image, you can line up the output with the original image. You can check if the face and clothes haven't changed, if the specified poses or equipment have appeared, and if the shape of the close-up parts hasn't collapsed. When using regular H3 instructions in Turbo, it is an application to a different model like in example 13. You can start as one piece to compare input and results, rather than on the premise that the same thing as the original work will come out. Before making a game or program app, see what can be done with a single image and public instructions. Among these 15 cases, there are works where you can trace up to that entry point.
