Kling 2.5 is a freak

Key Concepts

Cling 2.5 Turbo, AI video generation, text-to-video, image-to-video, physics understanding, anatomy, dynamic camera movements, prompt adherence, Hyo O2, VO3, cinematic quality, multi-elements feature, Chat LLM, Abacus AI, credits, audio synchronization.

Cling 2.5 Turbo: Overview and Capabilities

Cling 2.5 Turbo is a new and improved AI video model that excels in generating high-action scenes and demonstrating a strong understanding of physics. It is significantly cheaper (30x) than Cling 2.1, requiring fewer credits per generation. The model supports both text-to-video and image-to-video generation.

Key improvements:

• Physics and Anatomy: Demonstrates superior understanding of physics and anatomy compared to previous models and competitors.
• Dynamic Camera Movements: Excels at understanding and generating dynamic camera movements specified in prompts.
• Prompt Adherence: Accurately interprets and executes complex prompts with multiple elements.
• Cost-Effectiveness: Significantly cheaper to use than previous versions.

Official Demos and Examples

The video showcases several official demos highlighting Cling 2.5 Turbo's capabilities:

• Spacecraft Scene: A gray spacecraft racing through an asteroid belt with blue thrusters, demonstrating dynamic low-angle tracking and fast zooms.
• Ruined City Scene: A wide shot of a ruined city with collapsed towers and fires, followed by a fast camera drop and zoom to a person's eyes reflecting an approaching enemy, then zooming out to reveal a flaming halo. This showcases complex camera perspectives and prompt understanding.
• Basketball Game: A front tracking shot of a basketball game.
• Creepy Clown: A close-up of a clown laughing, then fading with black tears, followed by a sinister grin. This demonstrates the model's ability to generate expressions and emotions, including audio.
• Image-to-Video Examples:
  • A wand waving with lightning crackling around it, forming a translucent stag.
  • Dancing figures with a 360° top-down shot, preserving character consistency and minimizing warping.

Personal Demos and Comparisons with Hyo O2 and VO3

The presenter conducts personal demos comparing Cling 2.5 Turbo with Hyo O2 and Google's VO3 using the same prompts.

Key Findings:

• Gymnast on Balance Beam: Cling 2.5 Turbo generates the most anatomically accurate gymnast performing a flip, with accurate mirror reflection. VO3 performs poorly, while Hyo O2 is decent but not as realistic.
• Breakdancing Performance: Cling 2.5 Turbo excels at generating a breakdancing scene with flips and intense spins, including aggressive camera movements with speed ramps and whip-like acceleration/deceleration. VO3 and Hyo O2 have numerous errors.
• Princess and Dragon: Cling 2.5 Turbo and Hyo O2 perform well, but Cling's generation is preferred. VO3's scene is flawed with a sudden dragon appearance and slow movement.
• Confused Man in Marketplace: Cling 2.5 Turbo accurately generates a close-up of a confused man in a crowded marketplace with explosions, followed by a zoom out and rotation. Hyo O2 is decent, while VO3 is poor.
• Parkour Athlete: Cling 2.5 Turbo handles a parkour scene with flips and a first-person tracking shot best. Hyo O2 is decent in physics but lacks the tracking shot.
• Restaurant Kitchen: Cling 2.5 Turbo generates a time-lapse orbit shot of a busy restaurant kitchen. Hyo O2 is also good, while VO3 fails to generate an orbit shot.
• Man Riding Unicycle and Juggling: None of the models get it completely correct, but Cling 2.5 Turbo generates the most realistic unicycle riding, with fewer errors in juggling.
• Snowboarder Launching off Cliff: Cling 2.5 Turbo generates a smooth, cinematic, and physically accurate scene. Hyo O2 has slight physical errors, while VO3 fails completely.
• Kung Fu Fight Scene: Cling 2.5 Turbo handles fight scenes well. Hyo O2 is decent but with lower quality. VO3 fails to generate a decent fight scene.
• Ballerina in Studio: Hyo O2 nails the prompt, including the elephant balancing on a circus ball. Cling 2.5 Turbo generates the ballerina spinning and includes all elements, but the elephant is frozen. VO3 performs poorly.
• Couple Kissing with Text "The End": Hyo O2 nails the prompt, including camera movements and text. Cling 2.5 Turbo generates random text. VO3 generates the text but with incorrect camera movements.
• Professor Writing "Hello" on Chalkboard: Cling 2.5 Turbo starts with "H" and then writes gibberish. Hyo O2 and VO3 are closer but still incorrect.
• Will Smith Eating Spaghetti: Cling 2.5 Turbo generates a person with a similar haircut but not Will Smith's face. Hyo O2 generates Will Smith accurately. Using image-to-video with a photo of Will Smith, Cling 2.5 Turbo and Hyo O2 perform well, while VO3's result is worse.
• SUV Driving on Bumpy Road: Cling 2.5 Turbo preserves the character's look and follows the prompt. Hyo O2 goes overboard with the jiggle physics. VO3's quality and consistency are lower.
• Busy Marketplace: Cling 2.5 Turbo has the fewest errors and best consistency. VO3 has noise and errors, especially in the background. Hyo O2 has more noticeable errors than Cling 2.5 Turbo.
• Dancing in Sync: Cling 2.5 Turbo and Hyo O2 are good, but Cling 2.5 Turbo preserves faces better. Hyo O2 has more action but less consistent faces. VO3 is a mess.
• Anime Girl Talking to Creature: Cling 2.5 Turbo nails the prompt with a slow zoom in. Hyo O2 zooms in too quickly. VO3 doesn't zoom in or have the girl talking.

Chat LLM by Abacus AI

The video is sponsored by Chat LLM by Abacus AI, an all-in-one platform for using various AI models, image generators, and video generators. It offers features like seamless model switching, an artifacts feature for previewing generations, and a deep agent feature for complex tasks. The platform costs $10 a month.

Using Cling 2.5 Turbo

Cling 2.5 Turbo is available on Cling's online platform. Users can select the 2.5 Turbo model and choose between text-to-video and image-to-video generation. The platform offers pre-built prompts for controlling camera movement and speed. A 5-second video costs 25 credits, while a 10-second video costs 50 credits. The platform also has an option to add sound to videos, but the audio synchronization is not as good as VO3. The multi-elements feature is not yet available with the 2.5 Turbo version.

Benchmark Scores

Cling 2.5 Turbo outperforms Seed Dance 1.0 and VO3 fast in text-to-video and image-to-video benchmarks. However, the benchmarks did not include Hyo O2, which is a strong competitor.

Conclusion

Cling 2.5 Turbo is a leading AI video model that excels in physics and anatomy understanding, generating dynamic camera movements, and creating high-action scenes. It is a significant improvement over previous versions and a strong competitor to other leading models like Hyo O2 and VO3. While it has some limitations, such as generating text within videos and audio synchronization, its strengths make it a valuable tool for AI video generation.