Best in class models. Final 2025 snapshot.
Coding belongs to Cursor, Anthropic with Opus and Sonnet 4.5 leading real production work.
Images are dominated by Nano Banana Pro, Seedream 4.5, and GPT-Image 1.5.
Video is a three-horse race. Google Veo 3.1 for realism, OpenAI Sora 2 for creativity, Kling 2.6 for control.
Planning and reasoning clearly goes to GPT-5.2 Thinking.
Research speed king is Google Gemini 3.0 Flash.
Real time news is owned by xAI Grok 4.1.
Big winners this year. Google, Anthropic, and Grok aggressively gained market share.
Big losers. Midjourney and Perplexity AI faded hard from mainstream relevance.
Coding belongs to Cursor, Anthropic with Opus and Sonnet 4.5 leading real production work.
Images are dominated by Nano Banana Pro, Seedream 4.5, and GPT-Image 1.5.
Video is a three-horse race. Google Veo 3.1 for realism, OpenAI Sora 2 for creativity, Kling 2.6 for control.
Planning and reasoning clearly goes to GPT-5.2 Thinking.
Research speed king is Google Gemini 3.0 Flash.
Real time news is owned by xAI Grok 4.1.
Big winners this year. Google, Anthropic, and Grok aggressively gained market share.
Big losers. Midjourney and Perplexity AI faded hard from mainstream relevance.
👍439🔥146🎉140❤136
I tested YouArt, an autonomous AI video agent that turns a single chat prompt into a fully finished multi scene video. No templates, no timelines, no manual stitching. You describe the idea, the agent plans the scenes, generates visuals, handles pacing, transitions, and delivers a ready to post video in one workspace. I pushed it through emotional short films, Pixar style animation, cinematic Lego visuals, global perfume ads, and TikTok style UGC. Same flow every time, one prompt in, finished video out, with full transparency so you can inspect and tweak every step. If you are into AI video, autonomous agents, and next level workflows, this feels like Cursor for video.
https://youtu.be/hc_LcgsXSfk
https://youart.ai?utm_content=vortex
https://youtu.be/hc_LcgsXSfk
https://youart.ai?utm_content=vortex
YouTube
YouArt - This AI Talks Back With Finished Videos, Cursor But For Video
Join YouArt now https://youart.ai?utm_content=vortex
I tested YouArt, an autonomous AI video agent that turns a single chat prompt into a fully finished multi scene video. No templates, no timelines, no manual stitching. You describe the idea, the agent plans…
I tested YouArt, an autonomous AI video agent that turns a single chat prompt into a fully finished multi scene video. No templates, no timelines, no manual stitching. You describe the idea, the agent plans…
👍642🎉36❤26🔥25
This media is not supported in your browser
VIEW IN TELEGRAM
DreamID-V by ByteDance
It’s essentially like TikTok’s Face Fusion — face replacement in video.
The first diffusion transformer for high-quality face replacement in videos. It bridges the gap between image- and video-based approaches, delivering exceptional identity similarity and temporal consistency even in complex scenarios.
Demo: https://guoxu1233.github.io/DreamID-V/ Project: https://guoxu1233.github.io/DreamID-V/ Code: https://github.com/bytedance/DreamID-V
It’s essentially like TikTok’s Face Fusion — face replacement in video.
The first diffusion transformer for high-quality face replacement in videos. It bridges the gap between image- and video-based approaches, delivering exceptional identity similarity and temporal consistency even in complex scenarios.
Demo: https://guoxu1233.github.io/DreamID-V/ Project: https://guoxu1233.github.io/DreamID-V/ Code: https://github.com/bytedance/DreamID-V
👍15🔥6❤5🎉1
This media is not supported in your browser
VIEW IN TELEGRAM
I’ve seen several generations of 3D displays.
But what Samsung showed at CES 2026 looks pretty killer.
The thickness of the TV itself is especially impressive.
But what Samsung showed at CES 2026 looks pretty killer.
The thickness of the TV itself is especially impressive.
👍454❤159🎉148🔥138
Higgsfield has showcased a very serious relighting tool.
From some of the demos it was clear that it works extremely well with portraits, but then I also found this one where entire scenes are being relit!
It looks genuinely impressive. The available tools include selecting light direction, lighting setups, temperature, intensity, color, and shadow control.
Of course, you won’t be able to relight a whole scene for a film in a fully professional, exactly-the-way-you-want manner, but for low-budget production and advertising it’s more than good enough.
From some of the demos it was clear that it works extremely well with portraits, but then I also found this one where entire scenes are being relit!
It looks genuinely impressive. The available tools include selecting light direction, lighting setups, temperature, intensity, color, and shadow control.
Of course, you won’t be able to relight a whole scene for a film in a fully professional, exactly-the-way-you-want manner, but for low-budget production and advertising it’s more than good enough.
👍1.15K🎉398🔥373❤363
Qwen-Image-Edit-2511-Multiple-Angles-LoRA
An interesting tool for camera angles, equipped with a full ControlNet.
On the downside, the image quality isn’t great — the idea is cool, but the execution falls short.
https://huggingface.co/spaces/multimodalart/qwen-image-multiple-angles-3d-camera
An interesting tool for camera angles, equipped with a full ControlNet.
On the downside, the image quality isn’t great — the idea is cool, but the execution falls short.
https://huggingface.co/spaces/multimodalart/qwen-image-multiple-angles-3d-camera
👍648🔥246🎉239❤229
This media is not supported in your browser
VIEW IN TELEGRAM
Higgsfield “What’s Next?”
Higgsfield seem to be aiming to completely remove the traditional scripting component from content creation. That is, there will still be a “script,” but it will be written—or rather assembled—from AI-generated fragments. And not in text form, but directly as video snippets.
Higgs’s new feature, “What Happens Next,” lets you upload a SINGLE image, after which the AI suggests EIGHT video (!) variations of how the events could unfold. You choose the one you like, watch it to the end, and then once again pick one of eight possible continuations.
Higgsfield seem to be aiming to completely remove the traditional scripting component from content creation. That is, there will still be a “script,” but it will be written—or rather assembled—from AI-generated fragments. And not in text form, but directly as video snippets.
Higgs’s new feature, “What Happens Next,” lets you upload a SINGLE image, after which the AI suggests EIGHT video (!) variations of how the events could unfold. You choose the one you like, watch it to the end, and then once again pick one of eight possible continuations.
👍956🔥9❤8🎉8
GLM-Image
We’ve got a new open-source image generator, and technically it’s quite interesting. Earlier, Zhipu released the open-source LLM GLM, which crushed benchmarks and impressed many (you can try it at https://chat.z.ai/). Rumors of an image model followed — and now it’s here.
It’s already available on FAL: https://fal.ai/models/fal-ai/glm-image https://fal.ai/models/fal-ai/glm-image/image-to-image
The key idea is separating “thinking” from rendering. A 9B-parameter autoregressive model interprets complex, knowledge-heavy prompts, then passes them to a 7B-parameter diffusion decoder for rendering. With a custom Glyph Encoder, it aims to render text accurately inside images. Editing and style transfer are included out of the box. They claim quality on par with top diffusion models and better performance on complex tasks.
In practice, results so far look modest. Editing features need more testing and don’t seem very strong yet.
We’ve got a new open-source image generator, and technically it’s quite interesting. Earlier, Zhipu released the open-source LLM GLM, which crushed benchmarks and impressed many (you can try it at https://chat.z.ai/). Rumors of an image model followed — and now it’s here.
It’s already available on FAL: https://fal.ai/models/fal-ai/glm-image https://fal.ai/models/fal-ai/glm-image/image-to-image
The key idea is separating “thinking” from rendering. A 9B-parameter autoregressive model interprets complex, knowledge-heavy prompts, then passes them to a 7B-parameter diffusion decoder for rendering. With a custom Glyph Encoder, it aims to render text accurately inside images. Editing and style transfer are included out of the box. They claim quality on par with top diffusion models and better performance on complex tasks.
In practice, results so far look modest. Editing features need more testing and don’t seem very strong yet.
👍925🔥301❤276🎉273
This media is not supported in your browser
VIEW IN TELEGRAM
Hunyuan3D has been updated to version 3.1.
You need to take a look at the mesh, but it looks really polished.
Probably the most advanced 3D generator available today.
You need to take a look at the mesh, but it looks really polished.
Probably the most advanced 3D generator available today.
👍336🔥125🎉110❤100
This media is not supported in your browser
VIEW IN TELEGRAM
Wan 2.6 Image to Video Flash
So far, it works only from the first frame.
Video length: up to 15 seconds.
You can upload your own audio / audio generation is also available.
There is a shot_type option — single shot or multiple shots within one video.
Very fast.
https://fal.ai/models/wan/v2.6/image-to-video/flash https://wavespeed.ai/models/alibaba/wan-2.6/image-to-video-flash
So far, it works only from the first frame.
Video length: up to 15 seconds.
You can upload your own audio / audio generation is also available.
There is a shot_type option — single shot or multiple shots within one video.
Very fast.
https://fal.ai/models/wan/v2.6/image-to-video/flash https://wavespeed.ai/models/alibaba/wan-2.6/image-to-video-flash
👍385🔥129🎉125❤122
This media is not supported in your browser
VIEW IN TELEGRAM
Runway 4.5 Image to Video
A few days ago, Runway released an update. The main focus is on the Image-to-Video model. On their Twitter and website they show the best examples, but I took real generations and even found a comparison with Kling and Seedance.
I can’t say it’s some kind of revolution. The quality is not better than Kling. Length: 5–10 seconds. 720p.
A few days ago, Runway released an update. The main focus is on the Image-to-Video model. On their Twitter and website they show the best examples, but I took real generations and even found a comparison with Kling and Seedance.
I can’t say it’s some kind of revolution. The quality is not better than Kling. Length: 5–10 seconds. 720p.
👍370❤117🔥107🎉106
And a bit of rankings from Video LMArena.
Veo wipes the floor with everyone, especially in Text-to-Video.
In Image-to-Video, wan2.5 takes 3rd place, Seedance is 6th, and Kling 2.6 is 7th.
You can see that the amount of data is still pretty limited. Runway 4 is hanging out somewhere near the bottom, and for some reason mochi-1 from a year ago has snuck into the rankings.
But Veo’s hegemony will be very hard to beat.
LTX doesn’t show up in the charts at all.
https://lmarena.ai/ru/leaderboard/text-to-video https://lmarena.ai/ru/leaderboard/image-to-video
Veo wipes the floor with everyone, especially in Text-to-Video.
In Image-to-Video, wan2.5 takes 3rd place, Seedance is 6th, and Kling 2.6 is 7th.
You can see that the amount of data is still pretty limited. Runway 4 is hanging out somewhere near the bottom, and for some reason mochi-1 from a year ago has snuck into the rankings.
But Veo’s hegemony will be very hard to beat.
LTX doesn’t show up in the charts at all.
https://lmarena.ai/ru/leaderboard/text-to-video https://lmarena.ai/ru/leaderboard/image-to-video
👍942🔥291🎉282❤268