Skip to content
Date and source

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

Nov 16Mon
  1. WaveSpeedAI 官方博客69

    AI Video Character Consistency: A Single-Shot Workflow

    WaveSpeedAI's Dora shares an AI video character consistency shot workflow, using character reference kits, fixed identity details, and two-pass review to avoid character drift. Vidu 2.0 accepts 1-3 reference images, MiniMax H3 Highspeed accepts up to 9 reference images and 3 reference videos; when fixing drift, change only one input at a time.

    Why it matters: This workflow provides specific practices for character reference kits, fixed identity details, and two-pass review, and can be directly transferred to multi-shot AI video production.

Nov 14Sat
  1. Luma News44

    Recommended Native Audio AI Video Generators: Luma and Google Veo

    Native audio AI video generators can synchronously output dialogue, sound effects, and ambient sound in a single generation; Luma and Google Veo are representative platforms. Luma's Ray 3.2 supports multiple keyframes, HDR, EXR export, and Motion Transfer, and uses Uni-1 to maintain project context. Google Veo is positioned as an enterprise solution with broad platform access capabilities.

Nov 10Tue
  1. MiniMax · 音视频模型58

    MiniMax releases MiniMax-H3-Image-PE prompt expansion model

    MiniMax releases MiniMax-H3-Image-PE, fine-tuned from Qwen3.8-27B, providing prompt expansion and aspect ratio recommendations for MiniMax-H3-Image. It supports two workflows: text-to-image and image editing. The image editing PE is still under development; only its prompt is released this time. The model runs with FP4 weights and can be deployed via SGLang or Diffusers, requiring approximately 19 GiB of GPU memory.

Nov 9Mon
  1. InvokeAI 官方发布75

    InvokeAI 7.0.0 beta 1 released

    InvokeAI 7.0.0 enters beta 1, its largest version update. It adds projects, a single workbench, and video panel support for MiniMax H3, LTX-2.5, and Wan 2.2. Official warning: v7 database migration is irreversible; back up before upgrading.

    Why it matters: InvokeAI 7.0.0 enters beta, adding projects, a video panel, and multi-model support, giving creators a basis to evaluate the upgrade.

Nov 3Tue
  1. MiniMax · 音视频模型66

    MiniMax releases H3-Image, an open-source image generation and editing model

    MiniMax releases open-source image generation and editing model MiniMax-H3-Image, unifying text-to-image and image editing in a single checkpoint. The model is a 33.1B-parameter DiT, distilled to 6-step sampling, directly generating approximately 16-megapixel images (4096×4096 to 6272×2688) without super-resolution.

    Why it matters: The new model unifies text generation and image editing in a single checkpoint, directly outputs 16-megapixel images with only 6-step sampling, allowing evaluation of its suitability for existing image generation workflows.

Oct 22Thu
Oct 21Wed
  1. FastVideo 视频模型 · 模型发布80

    FastVideo releases ComfyUI version model weights for FastH3 Omni

    FastVideo released ComfyUI-usable weights for the final version of FastH3 Omni, providing two diffusion models, BF16 and INT8 ConvRot, which can generate video and synchronized audio from prompts and ordered image, video, and audio references.

    Why it matters: FastH3 Omni releases ComfyUI-usable BF16 and INT8 weights and provides the way to connect reference images, video, and audio, making it easier for creators to evaluate local video production pipelines.

Oct 20Tue
Oct 16Fri
  1. Higgsfield 官方博客38

    7 best Veo alternatives: how to maintain consistency during generation

    The official Higgsfield blog compared 7 Veo alternative tools, focusing on the consistency of characters, clothing, and scene details across multiple shots. Cinema Studio 4.0 supports up to 50 reference assets, 30-second clips, and 1080p output, with a 10-second 720p clip including audio consuming 70 credits; Seedance 2.5 supports 30-second native stereo and 50 references, and Kling 3.0 can generate up to 6 shots at a time.

Oct 13Tue
  1. Descript · 官方产品更新22

    Descript October update: global search, asynchronous recording requests, and ElevenLabs v4 voice cloning

    Descript released its October update, adding a search bar and a full results page to the entire cloud drive, and supporting Underlord in finding files, folders, projects, and layout packs. Enterprise cloud drives can send asynchronous recording requests, letting guests record on their own schedule and read scripts with a teleprompter; AI Speech has been upgraded to ElevenLabs v4, making voice clones closer to the real person.

Oct 9Fri
  1. Comfy 官方博客74

    How I generate real-time video on a single GPU with MiniMax H3

    On a single RTX 5090 (32GB VRAM), the author used FastVideo's FastH3 V2 checkpoint with four-step sampling to produce 15-second 448×256 real-time video, aiming to bring down the cost of continuous generation from 6k per day.

    Why it matters: The author breaks down one by one the six optimizations for running real-time video on a single RTX 5090, which can be transferred to your own production workflow.

Oct 6Tue
  1. Hedra 官方博客60

    Hedra Launches Developer API: One Key to Access 110+ Generative Models

    Hedra launches its Developer API, allowing access to 110+ generative models with a single API key and prepaid USD wallet, covering video, image, audio, and text, including Veo 3.1, Kling V3, Seedance 2.5, MiniMax H3, Nano Banana Pro, ElevenLabs V3, and more.

    Why it matters: Hedra consolidates 110+ video, image, audio, and text models into a single API and prepaid wallet, enabling developers to evaluate the integration costs of multi-model pipelines.

  2. FastVideo 视频模型 · 模型发布62

    FastVideo Releases FastH3 Trim Lightweight Model and ComfyUI Repackaged Files

    FastVideo released the experimental lightweight model FastH3 Trim, reducing transformer blocks from 50 to 42, using rank-16 timestep conditioning, 8-step sampling, with a 4.2x smaller size than the base H3 and faster than FastH3 V2.

    Why it matters: FastH3 Trim offers a lighter video generation solution with 42 layers, 8-step sampling, and 4.2x size compression, along with ComfyUI files and parameters for various precisions.

Oct 3Sat
  1. FastVideo 视频模型 · 模型发布62

    FastVideo releases the FastH3 Trim 8-step audio-video generation model

    FastVideo released the experimental model FastH3 Trim, which removes 8 of the least impactful blocks from FastH3 V2's 50 MiniMax H3 transformer blocks and replaces each block's AdaLN timestep projection with a shared rank-16 basis, trained with 8-step DMD2 (checkpoint 300), enabling 8-step text-to-video generation with synchronized audio.

    Why it matters: FastH3 Trim uses pruning plus shared AdaLN to bring 8-step audio-video generation to consumer-grade hardware, and can be compared against V2 to weigh image quality against size.

Sep 25Fri
  1. Hedra 官方博客62

    Hedra Compares API Pricing Across 11 Video Models Including Seedance, Kling, Veo, and MiniMax: 10-Second 1080p with Audio Ranges from $0.60 to $7.91

    Hedra compared API pricing across 11 video models for the same task (10 seconds, 1080p, native audio, one starting frame plus a text prompt), with the lowest at $0.60 for LTX-2.3 Fast and the highest at $7.91 for Seedance 2.5—a 13x difference. Nine models meet the spec; Veo 3.1 maxes out at 8 seconds and MiniMax H3 has no 1080p setting, so both are listed separately as close substitutes.

    Why it matters: For the same 10-second 1080p with audio task, API unit prices across 11 video models differ by 13x, along with a reusable method for model selection and cost verification.

Sep 18Fri
  1. fal 官方博客62

    fal Details H3 Max: Built on the fal Inference and Training Platform

    fal's official blog explains how H3 Max completes the journey from post-training to distribution across its three-layer platform: Compute, Serverless, and Model API. H3 Max generates a 5-second video in under 3 seconds and ranked first in quality, prompt understanding, and aesthetics in fal's human preference evaluation; independent benchmarks from Artificial Analysis and Design Arena also ranked it first.

    Why it matters: fal officially breaks down the three-layer platform that takes H3 Max from training to launch, offering a look at the inference and scaling mechanisms behind low-latency video generation.

Jan 27Tue
  1. Hugging Face Blog22

    Architecture Choices in China's Open-Source AI Ecosystem: How MoE, Multimodality, and Domestic Hardware Can Surpass DeepSeek

    Chinese open-source models have nearly unanimously shifted to MoE architecture, with Kimi K2, MiniMax M2, and Qwen3 all adopting this approach, aiming for sustainable deployment at controlled costs. Open-source activity expanded to multimodality and agents starting February 2025, with StepFun's Step-Audio-R1.1 voice model outperforming closed-source models, and Tencent open-sourcing Hunyuan Video and Hunyuan 3D.