←Search results

Search “Video”

By relevance · Total 2288 results Clear search

Wednesday

  4. Story3 reports · 2 facts

    fal says FLUX 3 Max scored 1,139 Elo in third-party evaluation

    fal says FLUX 3 Max achieved 1,139 Elo on a third-party benchmark, entering the top tier of frontier image-to-video models based on independent human preference evaluations. The model was launched through a collaboration between fal Research and bfl_ai. Earlier reports said that through dedicated post-training and inference optimization, the model tied for first place in image-to-video quality on an independent third-party benchmark, generating a 10-second 720/768p video in about 12 seconds; other reports said it scored 1,139 on the Artificial Analysis image-to-video index, with generation speeds up to 22 times faster than Seedance 2.5. fal Research says it is continuing to post-train the model and will subsequently improve quality and performance.

    fal says FLUX 3 Max achieved 1,139 Elo on a third-party benchmark, entering the top tier of frontier image-to-video models based on independent human preference evaluations. The model was launched through a collaboration between fal Research and bfl ai. Earlier reports said that through dedicated post-training and inference optimization, the model tied for first place in image-to-video quality on an independent third-party benchmark, generating a 10-second 720/768p video in about 12 seconds; oth

    Matched fieldsSummary

Sunday

  1. Story1 reports · 1 facts

    Meta paper argues that proactive assistants fail at autonomously judging when to interrupt

    On October 21, 2026, Meta published a paper pointing out that the failure point of proactive assistants (such as AI glasses) lies in autonomously judging when to speak, not just what to say. The team built the ProPerMem benchmark, using first-person video paired with interference-rich full-year memories averaging 618K tokens, and featuring manually verified trigger windows. The paper reports that all baselines scored at most 15%, even frontier models using retrieval-augmented memory; when oracle evidence was provided within the window, GPT-5.3-Chat reached 96%, but when it was left to choose the timing itself, this dropped to 38%, and after adding retrieval it fell to 15%. Grasping both content and timing while reducing trigger frequency can raise accuracy from under 10% to 42.4%. The work is currently released as a paper, with no subsequent progress yet.

    On October 21, 2026, Meta published a paper pointing out that the failure point of proactive assistants (such as AI glasses) lies in autonomously judging when to speak, not just what to say. The team built the ProPerMem benchmark, using first-person video paired with interference-rich full-year memories averaging 618K tokens, and featuring manually verified trigger windows. The paper reports that all baselines scored at most 15%, even frontier models using retrieval-augmented memory; when oracle

    Matched fieldsSummary

  2. Story2 reports · 2 facts

    Viggle AI Launches JST-2 on Its API

    Viggle AI announced that the JST-2 model is now available on its API, and that the model can generate 3D characters, motions, and scenes within a single architecture. Previously, on October 20, 2026, Viggle released JST-2, building the entire model stack in 3D space and representing the world with Physics-Native Tokens that carry xyz coordinates, attribute vectors, and dynamics vectors, rather than implicitly inferring physical laws from video pixels; the architecture is divided into three parts: the JST encoder, the JST generator, and the JST decoder, and a single forward inference generates a complete 3D scene in about 1 second at a cost of less than $0.01 (B200), and supports up to 64 input views as a spatial context window. Blind human preference tests published by Viggle show that for character generation it achieves a 58.4% win rate against Meshy 7.1 and a 73.2% win rate against Tripo H3.1, and for video-generated motion it achieves an 89.1% win rate against DeepMotion. Viggle AI previously described JST-2 as a "full-stack physics-native world foundation model" that can generate 3D scenes in seconds and extend them in real time. JST-2 is currently available externally through the API.

    Viggle AI announced that the JST-2 model is now available on its API, and that the model can generate 3D characters, motions, and scenes within a single architecture. Previously, on October 20, 2026, Viggle released JST-2, building the entire model stack in 3D space and representing the world with Physics-Native Tokens that carry xyz coordinates, attribute vectors, and dynamics vectors, rather than implicitly inferring physical laws from video pixels; the architecture is divided into three parts

    Matched fieldsSummary

  3. Story1 reports · 1 facts

    PyTorch 2.15 brings FP8 tensor support to MPS

    On October 20, 2026, PyTorch 2.15 added support for torch.float8_e4m3fn tensors on the MPS backend. Reporting from the same source also mentions that this version makes MoE model decoding 5–6.5 times faster than 2.14: on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B improved from 9.4 tokens/s to 61.3 tokens/s, and LFM2.5-8B-A1B improved from 23.6 tokens/s to 123.6 tokens/s. The report states that Wan2.2's tiled VAE video decoding peak driver memory dropped from about 270 GiB to 8 GiB; the image corruption issue in Qwen-Image-2.1's VAE encoding on MPS was also fixed with a padding change, with encode-decode round-trip PSNR rising from 10.06 dB to 29.95 dB. This version requires macOS 15 or later and Python 3.11 or later. There are currently no subsequent independent developments.

    On October 20, 2026, PyTorch 2.15 added support for torch.float8 e4m3fn tensors on the MPS backend. Reporting from the same source also mentions that this version makes MoE model decoding 5–6.5 times faster than 2.14: on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B improved from 9.4 tokens/s to 61.3 tokens/s, and LFM2.5-8B-A1B improved from 23.6 tokens/s to 123.6 tokens/s. The report states that Wan2.2's tiled VAE video decoding peak driver memory dropped from about 270 GiB to 8 GiB; the imag

    Matched fieldsSummary

  4. Story1 reports · 1 facts

    grok bot released version 0.71.2

    grok bot released version 0.71.2. According to information published by the author Eric Zakariasson on October 20, 2026, this version has smarter notification features; a single message supports up to 20 files, with a file limit of 250 MB and a video limit of 1 GB; forms can be pre-filled. The author says there are more updates. There are currently no subsequent developments.

    Matched fieldsSummary

  5. Story1 reports · 1 facts

    Hao AI Lab announces the launch of FastH3-QAD-v2 and the H3 NVFP4 VAE decoder

    Hao AI Lab announced the launch of FastVideo FastH3-QAD-v2 and the H3 NVFP4 VAE decoder. The team said this 768p video was generated on two DGX Sparks using QAD-v2 and the two new decoders, with 8-step video and audio generation, and using QAD weights. Currently, this news is released by Hao AI Lab as the first-party source, and there are no follow-up developments or cross-verification from other independent sources.

    Matched fieldsSummary

Monday

  1. Story1 reports · 1 facts

    NVIDIA releases surgical generative simulation model Cosmos3-H-Dreams

    NVIDIA has released the surgical generative simulation model Cosmos3-H-Dreams. The model is specialized from the 16-billion-parameter Cosmos3-H-Embodied-Nano for action-conditioned video generation of cholecystectomy. Official information shows that the model was trained with 28 action dimensions, 832 x 480 video, and 10 Hz action alignment, and was distilled into two student checkpoints, four-step and two-step; the two-step version was trained separately and offers faster inference but lower image quality. NVIDIA also provides the FlashDreams streaming runtime, as well as methods for adapting the model to proprietary data and controllers, and states that the model is a research simulator, not a validated clinical system. There are currently no subsequent development facts.

    NVIDIA has released the surgical generative simulation model Cosmos3-H-Dreams. The model is specialized from the 16-billion-parameter Cosmos3-H-Embodied-Nano for action-conditioned video generation of cholecystectomy. Official information shows that the model was trained with 28 action dimensions, 832 x 480 video, and 10 Hz action alignment, and was distilled into two student checkpoints, four-step and two-step; the two-step version was trained separately and offers faster inference but lower im

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    Generalist demonstrates GEN-1.5 organizing hockey pucks in a container

    On October 20, 2026, Generalist released a video showing its model GEN-1.5 loading 12 hockey pucks into a container; when a puck gets stuck or is in the wrong position, the model adjusts its placement. Generalist said more details are available in a blog post and provided a waitlist link in the comments. There are currently no follow-up developments supported by evidence.

    Matched fieldsSummary

Tuesday

Saturday