Skip to content
Date and source

Dates follow the news timeline in Beijing time and include both endpoints.

Nov 17Tue
  1. Reddit comfyui · 创作与工作流分享55

    Single RTX 3090 Real-World Test: Daily Usage Data for Local Agents and Image/Video Generation

    The author used a single RTX 3090 24GB, swapping VRAM in and out, to run a local agent and ComfyUI image/video generation simultaneously. Measured: LLM about 72 tok/s, SDXL image about 32 seconds, Wan 2.2 video 4-6 minutes. Data comes from 205 real requests and 35 rendering tasks, not synthetic benchmarks. The author concludes that 24GB can only accommodate either the LLM or one render at a time, and recommends using parallel 1 instead of 4.

Nov 16Mon
  1. Reddit StableDiffusion · 创作者与工作流分享34

    How to get the highest quality video from Minimax H3 on a 4090? What about upscaling?

    Minimax H3 on a 4090 achieves best quality with settings of 1322x768, 24fps, 20 sampling steps, without turbo and without LoRA; the duration is usually about 4 seconds. For upscaling to 2k, you can use Regenerate 2k (Regen2k IC just released) or minimax_h3_latent_upscaler_3d_conv_v1_fp16. The author tested the API at 1080p and believes the quality improvement is not worth the price.

Nov 13Fri
  1. Reddit comfyui · 创作与工作流分享61

    WakeGPU: Free open-source desktop app for remotely running ComfyUI/Colab GPU, with automatic idle shutdown and batch queuing

    WakeGPU is a free and open-source desktop client (macOS and Windows) that turns remote cloud instances into an automated private studio, supporting automatic idle shutdown and batch queuing. It is 100% free, MIT licensed, all advanced features work out of the box, no paywall, subscription, or telemetry, and binaries are released on GitHub Releases.

Nov 12Thu
Nov 11Wed
Nov 6Fri
  1. World Labs · 官方研究与公告60

    World Labs releases Atlas real-time research preview, Frameboy supports 15fps streaming exploration

    World Labs releases Atlas real-time research preview, demonstrating via Frameboy streaming frame generation from cloud GPU to browser at up to 15fps. This version highly optimizes Atlas while maintaining novel view synthesis capability; it is also integrated into Viewfinder, allowing users to build scenes frame by frame and then explore in real time.

    Why it matters: Real-time streaming generation reduces exploration cost to near zero, enabling creators to assess the potential of interactive world models to replace existing workflows.

Nov 1Sun
  1. Reddit StableDiffusion · 创作者与工作流分享77

    Author creates 46-second sci-fi trailer entirely locally on MacBook M5 Max and reveals complete pipeline

    Author completes 46-second sci-fi trailer 'THE LAST SIGNAL' entirely locally on a MacBook Pro M5 Max 128GB, with no cloud rendering and no API calls, taking about 14.5 hours (including about 2 hours of waiting).

    Why it matters: The author reveals the complete local pipeline from storyboard to final film and the actual time taken, allowing creators to assess the feasibility of replicating a similar workflow on a MacBook.

Oct 31Sat
Oct 30Fri
  1. Reddit comfyui · 创作与工作流分享66

    Kinoloom: An Open-Source Film Studio Built on ComfyUI, Supporting LTX 2.5 Seamless Long Takes

    Kinoloom is an open-source film studio based on ComfyUI that supports LTX 2.5 seamless long-take rendering and can run locally on an 8GB GPU. It offers a video studio, an image studio, and a film mode, can use any LLM to write stories, and supports MiniMax H3 (you need to supply your own weights due to licensing restrictions).

    Same story, the pick is 《Open-Source Local Video Tool Kinoloom Released, Supporting 8GB VRAM and Latent-Space Long-Take Stitching》

Oct 29Thu
  1. Loomance · 本地AI动画实测63

    Wan 2.2 Wild Mode Hands-on: How to Turn a Pencil Sketch into a Local Anime Video

    Wan 2.2 Wild Mode turns a pencil sketch into anime on a local PC. After hands-on testing, the author provides a two-step recipe: Wild selects the style, and Stable Scene controls the motion. Under the same prompt, five seeds produce five styles, four of which are usable; Wild automatically chooses camera and composition, while Stable Scene continues from a selected frame to lock the style and let the character perform.

  2. Reddit StableDiffusion · 创作者与工作流分享68

    Create a complete music video locally with LongCat-Video-Avatar 1.5 on a 5090, with Sync 3 lip-sync correction to within 3ms.

    The author used LongCat-Video-Avatar 1.5 on a local RTX 5090 to create a 3-minute music video and used Sync 3 for lip sync. The workflow uses INT8 DiT + bf16 distilled LoRA, 8 steps, SageAttention, T5 offloaded to CPU, 480p 25fps upscaled to 1080p.

    Why it matters: The author measured Sync 3 lip-sync output leading by 63-87ms; after per-shot correction with SyncNet, the error dropped to within 3ms. The method can be transferred to other lip-sync workflows.

Oct 25Sun
  1. TechCrunch · AI60

    Block releases speech model Dessa, which can run locally on consumer-grade GPUs

    Block, the research lab under Jack Dorsey, has released the speech model Dessa. The model is trained entirely on human speech, can run locally on a single NVIDIA GPU, and can speak and listen at the same time. Block calls Dessa a full-duplex speech-to-speech model, with expressions such as emotion, pauses, and laughter built into the model, and it can run on a MacBook M4 Max or higher-spec consumer computer. Dessa is not yet publicly available, and Block is inviting security researchers to participate in testing.

Oct 24Sat
  1. Sports Video Group34

    Prime Video debuts AI virtual camera system Prime 360 in its second NBA season

    Prime Video begins its second NBA season of coverage, producing more than 100 regular-season and playoff games this season, all in native 1080p HDR with 5.1 surround sound. The core new technology is Prime 360, an AI virtual camera system based on AWS Trainium chips that reconstructs a 3D scene from high-resolution cameras around the court to generate cinematic replay angles that traditional camera positions cannot capture.

Oct 23Fri
Oct 22Thu
  1. Hugging Face Daily Papers38

    DirCast: Esports Live Directing and Commentary with a Single Streaming Video Language Model

    DirCast uses a single streaming video language model to simultaneously perform directing and commentary for live esports broadcasts: it scans all player camera feeds at low resolution every second, a hold-or-switch directing module decides whether to cut the camera, and then it reads the selected camera feed at high resolution before speaking commentary.

Oct 21Wed
  1. AI资源分享站 · 开源创作项目分享58

    Open-source web-based video editing tool ai-video-editor released, supports local AI inference

    ai-video-editor is a browser-based, local-first open-source video editor that combines a multi-track timeline with on-device AI inference, completing editing, subtitles, and matting locally. The tool integrates Whisper small for audio-to-text with waveform-aligned subtitles, uses MODNet and YOLOS for portrait matting and subject avoidance, and includes MI-GAN for removing watermarks and clutter plus NanoVSR for 4× upscaling.

  2. Loomance · 本地AI动画实测69

    Wan 2.2 comparative evaluation under Loomance and WanGP: Fast Mode and full workflow yield different results.

    The author compared the rendering results of Wan 2.2 under two tools, Loomance and WanGP, and found that the 4-step lightning mode is about 10 times faster but has more errors, while the 30-step full sampling is cleaner. In the six-dog scene, WanGP had 3/5 instances of character melting or merging, while Loomance had 0/5; in high-resolution single-character scenes, the two are close. Choose as needed: use WanGP for rapid changes or low VRAM, use full sampling to maintain character consistency.

    Why it matters: A practical comparison of the same Wan 2.2 model under two local tools offers a reference for creators to choose between speed and fidelity.

Oct 20Tue
Oct 19Mon
  1. IT之家38

    NPCTalk brings real-time, script-free voice dialogue to GTA 5 NPCs

    The open-source AI mod NPCTalk brings script-free real-time dialogue to GTA 5: players can talk to NPCs via microphone, and NPCs respond with AI-cloned original voices. The system runs on NVIDIA Nemotron ASR, Qwen 3.8 27B (Cerebras platform), Fish Audio Drama 3, and TypeSafe Jev working together, with push-to-talk latency of about 1.15 to 2.2 seconds.

Oct 18Sun
Oct 16Fri
  1. Reddit StableDiffusion · 创作者与工作流分享12

    When will there be AI video generation models built specifically for 24-32GB of VRAM?

    A Reddit user posted a discussion on why there are still no video generation models designed specifically for 24-32GB of VRAM. They noted that this VRAM range is the upper limit ordinary consumers are willing to invest in a gaming/work PC, and that currently only quantized models fit, often requiring compromises on the encoder LLM as well.

  2. postPerspective36

    Nvidia Editor's Day: New Windows RTX Spark computers unveiled

    Nvidia unveiled the RTX Spark line of laptops and desktops at Editor's Day in San Francisco, emphasizing local AI processing and agentic AI. The SparkN1x chip integrates a 20-core Grace CPU with a Blackwell GPU, offering up to 128GB of unified memory, 6,144 CUDA cores, and 1 petaflop of NVFP4 compute, roughly six times the performance of Qualcomm's fastest solution.

Oct 15Thu
  1. Creative Bloq42

    I used Nvidia RTX Spark's AI agent to operate Blender, and here's how it went

    On an ASUS ProArt P16, the locally running Hermes agent (Qwen3.8 27B) could drive Blender to complete simple tasks such as warming the lighting, changing hair color, and exporting before-and-after comparison images, but struggled with complex multi-step edits: adding clown makeup to the model took about half an hour and only produced an approximate effect; moving an object across scenes, estimated to take 5-10 minutes after 25 minutes of preparation, was still unfinished after an hour and a half.

  2. 机器之心SOTA模型 · 公开公众号 RSS22

    StepFun opens Step 5 Preview weights, Ant Ling open-sources Ling-3.1-flash

    StepFun opened the official weights of Step 5 Preview, with 600B total parameters, 27B activated, support for a 1M context, and the ability to handle text and image understanding, coding, and long-horizon tool calling, providing BF16 and FP8 weights and vLLM and SGLang deployment paths. Ant Ling concurrently open-sourced Ling-3.1-flash under the MIT license, with 560B total parameters, 25B activated, support for a 1M context, aimed at office, search, and coding agent scenarios, and providing a hybrid reasoning switch.

  3. note · loyal_owl5720 · 图像视频创作分享22

    AMD XDNA 2 NPU speeds up Anima image generation by about 40%, bringing a single image down to the 35-second range

    When running Anima image generation on the AMD XDNA 2 NPU, by adjusting the timing of the MIMALLOC_PURGE_DELAY=0 setting, generation speed improved from about 60 seconds to an average of 35.81 seconds per image, a speedup of about 40%. This optimization only takes effect when the NPU is used; generating on the CPU alone still takes an average of 55.47 seconds per image. It is currently unclear why this setting is only effective for the NPU, and power consumption will be quantitatively evaluated later.

  4. IT之家50

    NVIDIA Broadcast 2.3.0 released: compatible with RTX Spark, experimental support for video super resolution and frame generation

    NVIDIA Broadcast 2.3.0 was released, adding support for RTX Spark devices and introducing two AI effects, RTX video super resolution and frame generation, in beta form. This version can apply supported effects to local media files, supports connecting to AI agents via MCP Server, and the audio studio effects offer four professionally tuned microphone profiles.

  5. IT之家62

    Kunlun Wanwei releases Matrix-Game 4.0 video large model game and open-sources it simultaneously

    Kunlun Wanwei released Matrix-Game 4.0, calling it the world's first video large model game running on a single consumer-grade graphics card, and has open-sourced it simultaneously. The model uses a game engine to maintain world state and interaction rules, with a video generation model rendering visuals in real-time, achieving 480p, 60 FPS on a single NVIDIA RTX 5090, with key-to-screen response latency under 150 milliseconds.

  6. No Film School38

    Shootsolo 2.0 released: voice-controlled camera app supports simultaneous landscape and portrait recording

    The voice-controlled camera app Shootsolo has released version 2.0, supporting saving both portrait and landscape videos in a single shoot, and adding a built-in teleprompter, virtual studio lighting, and background effects. This version removes the original points system, has no ads, no watermarks, and no subscription, can be used fully offline, and is available for free on the App Store and Google Play.

Oct 14Wed