ICN | B.IA.seed Edge (@icryptan) header
ICN | B.IA.seed Edge (@icryptan) profile picture

Get some edge and beat the market Building Animora, an AI video studio. Connecting creators, sharing discoveries & team lessons. AI video cheat sheet pinned ↓

web3
Joined October 2021
Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)Photo by ICN | B.IA.seed Edge (@icryptan)
AI video creator cheat sheet: compare the cost of 10 seconds across model makers and aggregators, find free credits, contests & creator programs with prizes and deadlines. Sources linked. Updated monthly. Free for followers. Enter your handle: yst.media/en/x
2
200
How to tell which model made a video, part 4 of 4 Last model is Grok Imagine 1.5 Here is what to use each of the 11 models for.
1
1
84
What to use each model for Kling 4.0 handles dialogue and takes up to 30 seconds based on keyframes. Seedance 2.5 is for that "looks like real footage" feel and for action, if the budget allows. Gemini Omni 1.2 Flash is for ads and editing an existing video with words. MiniMax H3 gets text and branding right in the shot and runs locally. PixVerse V6 Max makes a quick 1080p video from a single prompt. Vidu Q4 Preview is for acting, anime and a character's voice from references. Luma Ray3.2 is for restyling live-action footage and HDR for color grading. LTX-2.5 is for your own pipeline, speed and 50 frames per second. Wan 3.0 makes long dialogue scenes from many references. Runway Gen-4.5 is for silent, staged shots to cut into an edit. Grok Imagine 1.5 is a cheap way to animate photos with speech, including in Russian. The main trick is to check the label and file format first (length, frame rate, whether there's audio), then look at the image. The label and format are harder to fake than the style, and that usually narrows eleven candidates down to two or three. Sources were developer documentation, Curious Refuge, Genativity, TechPilot, VdoBloom, Megaton, GenVerdict, Promptowy, DataCamp, Quantslant, CNMO, 36Kr, The Paper, iFanr, AITopTools, Pixel Dojo, Android Central, Reddit and X.
1
58
Riddle Which model from the review made this clip? All the clues you need are in the breakdowns above. Post your guesses in the comments. The answer comes tomorrow.
35
How to recognize a model from a clip, part 3 of 4 Luma Ray3.2, LTX-2.5, Wan 3.0, and Runway Gen-4.5
1
1
1
143
Wan 3.0 For long dialogue scenes up to 30 seconds and for videos built from many references: images, video, audio and even documents. Through the API it's $1 per 10 seconds of 720p Label and format. In the Model Studio interface the watermark can't be turned off; in Alibaba's example it's an "AI generated" mark at the bottom right. Through the API there is no mark by default. 30 frames per second, while the other models with an open spec have 24 or 25 (LTX also has 48 and 50). Up to 30 seconds, up to 1080p. Picture and motion. Convincing facial expressions, but the skin is too detailed (Curious Refuge). On pans the camera jerks every five frames, and the shot unexpectedly changes through a semi-transparent dissolve (Reddit). Sound. Dialogue, music and effects in one pass; in the jump rope test the author liked the sound more than Omni and H3. Alibaba itself admits that the timbre and texture of the sound are not yet perfect. Typical failure. Hands: fingers all open at once, the sliced pepper is whole again (Genativity). Characters freeze like mannequins between actions and look into the camera. What they say. Alibaba promises to keep characters, objects and style consistent between scenes. On X, a production author describes a "mannequin effect" and a slow pace that gets sped up by 10-30% in editing .
1
64
Runway Gen-4.5 For short staged shots with precise camera control and a professional pipeline: HDR, ProRes, EXR. Via the API it's $1.20 for 10 seconds, on the Standard subscription $2.88. Label and format. Runway embeds C2PA metadata in the results of its models; on the free plan there is also a visible watermark. A silent clip, 2 to 10 seconds at 720p: Gen-4.5 does not generate sound, and speech is added by a separate model, Act-Two. In the Promptowy test, neither file had an audio track. Picture and motion. Cinematic picture and real camera pull-backs; on Reddit people note that objects often sit exactly in the center of the frame. Sound. It has no sound of its own. If a Runway clip has a voice, it was added with another tool. Typical failure. Physics: in V-Benchmark Megaton it scores 38.82 out of 100, text 47.22. The model skips the key action (the dog is already holding the frisbee) and, by Runway's own account, shows a successful outcome even where it's impossible. What people say. Runway calls 4.5 its most advanced generation model; the only newer things from it are the Aleph 2.0 editing tools and Ruby for HDR. On X it's called "the director's choice" for the timeline and character consistency.
69
How to tell which model made a video · part 2 of 4
2
1
141
PixVerse V6 Max For short ad and stylized clips in 1080p with sound from a single prompt. The cheat sheet has the price of the base V6: $0.40 per 10 seconds of 720p without sound and $0.53 with sound; V6 Max hasn't been priced yet. PixVerse via my affiliate link: motivaiprivatelimited.sjv.io… Label and format. On the Basic plan, export comes with a watermark. In the Stork test it's a PixVerse.ai inscription at the top right. With Standard the mark is removed. Up to 15 seconds and 1080p, eight aspect ratios. Unlike regular V6, Max takes only text or an image, with no first and last frame and no references. Picture and motion. Sharp, clear picture and a confident camera: in a comparison on Reddit, Max chose a wide shot and fleshed out the crowd, while V6 shot it as if from a phone inside the crowd. Sound. Dialogue, effects and background in one pass. In the Reddit test, two women got male robotic voices until the gender was specified explicitly. Typical failure. Background text is unreadable, objects hang in the air, and instead of pulling the camera back the model cuts (GenVerdict, Stork). With complex camera moves the frame corners flicker. What people say. PixVerse admits that precise control of direction in complex scenes is still weak. GenVerdict gave base V6 7.1 out of 10. On X people praise multi-layered clips and presets, for example a 90s anime feel (per a Grok summary). The review is based on V6, on which Max is built. #PixVerse #PixVerseV6Max
1
1
83
Vidu Q4 Preview For acting scenes, anime and a hero who has to stay consistent from images and voice. Through the API it is $0.89 for 10 seconds of 720p, priced in yuan, and this is a preview. Label and format. A visible "Created by AI" mark in one of the corners is switched on only if you want it. In the metadata Vidu writes its own label unless the user passes a different one. From 3 to 16 seconds, 24 frames, up to 4K in 10-bit. There is no text-only generation in the preview: only from an image or from references. Image and motion. Sweeping camera, fights and dense special effects: iFanr directly calls this "Vidu's manner". Fire lights up the characters, water changes the reflections to follow the angle. Sound. It clones a voice from three references and does lip sync. In The Paper's test the model got confused and added characters in a Chinese tongue twister, where Seedance 2.5 and Kling 4.0 handled it. Typical failure. Text turns into symbols, small details get duplicated: flying glasses grew an extra temple arm, Newton's cradle went chaotic after five seconds (The Paper). Distant faces are unreadable unless you specify a medium shot. What people say. Shengshu CEO Yihang Luo himself advises checking the model outside curated demos and admits that a usable shot often takes several attempts. Authors on X praise the expressive eyes and acting. #Vidu #ViduQ4
1
1
101
How to tell the model from the clip: signs of 11 video models, part 1 of 4 From a finished clip you can narrow it down to two or three models in a minute, and sometimes name it exactly. I'm going through Kling 4.0, Seedance 2.5, Gemini Omni 1.2 Flash, MiniMax H3, PixVerse V6 Max, Vidu Q4 Preview, Luma Ray3.2, LTX-2.5, Wan 3.0, Runway Gen-4.5 and Grok Imagine 1.5. Each breakdown covers: what the model is for, the price per 10 seconds, the signs from the most reliable to the most debatable, and what the developer and authors on X and Reddit say. Under each model there is an official clip and a frame. Prices are from my AI video creator cheat sheet: yst.media/ru/threads. Checked on November 2.
5
2
294
Seedance 2.5 For videos that have to look like real footage, and for action. Through the BytePlus API it's $2.31 for 10 seconds of 720p, and at 1080p it's already $5.69. Label and format. Through the API, the visible "AI生成" mark at the bottom right appears only with the parameter watermark=true. By default it is not there. In Dreamina, ByteDance promises an invisible mark, C2PA and visible AI labels. Up to 30 seconds per pass, 24 frames per second, 1080p in 10 bit, stereo sound on by default. Picture and motion. The texture of real footage: natural depth of field, subtle facial expressions, a rhythmic camera (36Kr). Curious Refuge, in six pairs against Kling 4.0, gives Seedance the edge in dynamics and physics. Sound. It adds a musical intro on its own, plus a rising sound at the end, even when music was forbidden (Reddit). Lip sync is reliable in medium and close shots. In wide shots the characters mumble. Typical failure. Gazes and positions of the characters jump between shots. In dark scenes people on Reddit see dirty gray spots. On pause, mirrored digits were found on the cards in the characters' hands, which you can't see in normal viewing (AITopTools). What people say. ByteDance itself admits that the physics of complex movements and scenes with several characters still need to be improved. Authors on X praise the clean sound without the "underwater hiss" and 30-second takes with no visible seams.
1
1
89
Gemini Omni 1.2 Flash For ads, try-ons and fixing a finished clip with words: light, background or camera angle without a new generation. The cheat sheet still has the price of the previous version, 1.1: $1.01 for 10 seconds of 720p with sound in the Gemini API. I haven't checked the 1.2 price yet. Label and format. Invisible SynthID and C2PA on everything created in Gemini, Flow and YouTube. This is the most reliable check in the review: upload a file to Gemini and ask whether it was made with Google AI tools (up to 100 MB and 90 seconds). From 3 to 15 seconds, 24 frames per second, up to 4K, native sound; with extensions up to 30 seconds. Picture and motion. Continuous fly-arounds with no cuts: in Quantslant's test it was the only one of six models that circled the saleswoman and came back to the original angle. Sound. Dialogues of several characters with lip sync, but in DataCamp's test the bell sounded 0.15 seconds after the hit, and after two edits the lag grew to 0.23. Typical failure. Objects disappear behind an obstacle and reappear without a transition, small details get swapped: Google itself showed how a ladle turned into chopsticks. In CNMO's test, out of ten generations only one kept the toy character fully intact. What people say. Google admits weak spots in fast sports and complex physics. On Reddit people write that 1.2 filters requests more strictly and sometimes follows the prompt worse than 1.1; Artificial Analysis on X rates it high in editing but weak in live footage with moving people (per a Grok summary).
94