Skip to content
Date and source

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

Nov 4Wed
  1. ComfyUI Wiki · 工作流与开源创作分享63

    PrunaAI releases the P-H3-Image image model, fine-tuned from MiniMax H3

    PrunaAI releases the P-H3-Image image model, fine-tuned from MiniMax H3, supporting 8 MP text-to-image generation and editing with up to 8 input images. The model is a 33.1B-parameter transformer, uses 8-step sampling without CFG, includes a built-in Qwen3.5-9B prompt upsampler, and is natively supported in ComfyUI v0.44.0.

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.

  2. ComfyUI Wiki · 工作流与开源创作分享68

    TaoLive AIGC releases TaoMate-FVSR, a video face restoration and visual dubbing model.

    On November 2, TaoLive AIGC released TaoMate-FVSR, a 10.3 GB video face restoration model that can fix blurry faces and fused teeth in generated videos, and make lip movements match new speech when replacing audio tracks.

    Why it matters: Addressing the defects of blurry faces, fused teeth, and lip movements not matching the audio in generated videos, it simultaneously performs restoration and lip-syncing, and can be run directly in ComfyUI.

Nov 2Mon
  1. 电影界 · 短剧经营与内容趋势63

    Tencent Hunyuan open-source simultaneous interpretation model Hy-SimulMT, real-time voice cloning assists in dubbing short dramas for overseas distribution

    Tencent open-sourced the Hunyuan simultaneous interpretation large model Hy-SimulMT, which supports real-time two-way translation between Chinese and English speech, and can clone the speaker's voice in real time while translating. Measured response is about 3 seconds, and translation quality reaches the level of human simultaneous interpretation experts. This capability is expected to reduce the cost of dubbing short dramas for overseas distribution, allowing multilingual versions to be replicated at low marginal cost, so small and medium-sized teams do not need to assemble a separate dubbing crew for each market.

Nov 1Sun
  1. Reddit StableDiffusion · 创作者与工作流分享42

    Morgana 1.2: FLUX.2 Klein 4B fine-tune for photography, realistic faces, natural light and fine textures (free weights + GGUF)

    Morgana 1.2 is released, a photography fine-tune model by Kitani based on FLUX.2 Klein 4B, focusing on realistic faces, natural light and fine textures. On 1x H100, 15-step generation takes only 4.6 seconds, no LoRA needed. Free safetensors and GGUF (including BF16, Q8_0 quantizations) are provided, with a paid API also available.

Oct 30Fri
  1. ElevenLabs 产品与研究60

    ElevenLabs launches SFX v3 sound effects model, now available on ElevenCreative and ElevenAPI

    ElevenLabs launches SFX v3 sound effects model, now available on ElevenCreative and ElevenAPI. SFX v3 handles prompts describing multiple sounds more reliably, produces more believable music and vocal generation, and delivers cleaner, more natural audio quality than SFX v2. Creators can use it in Flows, the Sound Effects tab, Studio, and Music Editor, while developers can integrate it via ElevenAPI.

    Why it matters: SFX v3 improves multi-sound prompt following, music and vocal generation, and audio quality, allowing creators to evaluate whether to replace their existing sound effects generation workflow.

  2. ComfyUI Wiki · 工作流与开源创作分享83

    Pika H3 Flex Release: Six-Step Video Generation Model with Open-Source Weights and ComfyUI Template

    Pika releases H3 Flex, which post-trains and distills MiniMax H3 to six-step sampling, with open-source weights and ComfyUI template. The model loads onto the H3 base with two LoRA adapters, supporting text-to-video, first/last frame, and reference-condition generation, outputting 480p/768p, 5-15 seconds, and mixing in 32kHz stereo audio. Pika's self-evaluation shows it leads comparison models on five visual and audio quality metrics; community testing suggests high-frequency detail flickers at four steps, recommending keeping six steps.

    Why it matters: Pika post-trains MiniMax H3 into a six-step model and open-sources the weights. The ComfyUI template can be loaded directly, allowing creators to evaluate the quality of six-step output accordingly.

  3. Pika68

    Pika H3 Flex released, topping the DesignArena t2v open-source weight leaderboard

    Pika releases H3 Flex, becoming the top open-source weight model on the DesignArena t2v leaderboard. The model supports exporting audio stems, up to 4K resolution, and faster-than-realtime generation; a 15-second clip with audio track can be generated in as fast as 7 seconds.

    Why it matters: Pika H3 Flex tops the DesignArena t2v leaderboard with open-source weights, supporting 4K and audio stems, providing a new option for professional iteration.

  4. GitHub:AI视频开源项目发现76

    Pika releases H3 Flex video generation model, supporting six-step sampling and multi-reference input.

    Pika releases H3 Flex video generation model, post-trained on MiniMax-H3, generating video with synchronized stereo audio via six-step sampling. Local open-source inference supports 480p/768p, 24fps, 5-15 seconds; API can generate 1080p/2K/4K and supports audio separation and prompt enhancement; Ref2VA mode supports up to 12 reference inputs.

    Why it matters: Post-trained on MiniMax-H3, six-step sampling and multi-reference input provide an efficient and controllable solution for professional video creation, and its compatibility with existing workflows can be evaluated.

Oct 29Thu
  1. Reddit StableDiffusion · 创作者与工作流分享73

    Krea 2 Turbo Distill 3-Step LoRA: 8 Steps Reduced to 3 Steps

    The creator released a 3-step distilled LoRA for Krea 2 Turbo, reducing denoising steps from 8 to 3, with denoising speed improved 2.6x and detail quality reaching 1.15x that of the teacher model. Training used 12,000 samples, about 41 hours on a single H100; a ComfyUI workflow is provided, compatible with diffusers, ComfyUI, and MLX.

    Why it matters: This 3-step distilled LoRA reduces Krea 2 Turbo's denoising steps from 8 to 3, with detail quality exceeding the teacher model, giving creators a faster image generation option.

  2. 可灵官方博客71

    Ami discusses how Kling 4.0 brings subtle emotions to life.

    Kling invited creator Ami to use Kling 4.0 to produce the short film "Tea When You Have Time". She shared her experience using new capabilities such as 30-second generation, Omni Reference, and multi-keyframe. Kling 4.0 supports single-run 30-second generation, 21:9 aspect ratio, up to 10 keyframes, consistent 12-character assets, good lip-sync for Cantonese dialogue, and supports multiple dialects and English accents.

    Why it matters: Creator hands-on testing shows Kling 4.0 performs stably in 30-second long takes, 12-character consistency, and Cantonese lip-sync, providing a basis for evaluating an upgrade.

  3. Reddit StableDiffusion · 创作者与工作流分享56

    Retro Anime Style LoRA for MOSS

    The author trained a retro anime style LoRA, recreating the cel texture, film grain, and hand-painted background style of 80s/90s OVAs. Recommended weight is 0.7-1.0, trigger word is retro anime style illustration, and it can be downloaded at https://civitai.red/models/1970260/retro-anime-style.

  4. Reddit StableDiffusion · 创作者与工作流分享27

    Ultimate Digital Artwork v2 - MOSS lora

    MOSS lora v2 released. The author cleaned low-quality images from the dataset and re-annotated all images, no longer using converted Ideogram prompts. The new model shows significant improvements in color depth, prompt adherence, and high-detail image generation. Due to performance improvements, the model is less adventurous than v1, and prompts need to be more specific.

Oct 28Wed
  1. Reddit StableDiffusion · 创作者与工作流分享32

    AuroraIMG2-50M-Preview released!

    AuroraIMG2-50M-Preview has been released. It is a text-to-image model with a denoiser of only 50M parameters and approximately 90M end-to-end inference parameters. On a 2016 Intel i5 7200U laptop CPU, generating with 50 steps takes 27 seconds, and with 20 steps takes 11 seconds. It is released under the Apache 2.0 license.

  2. 智东西65

    Google releases next-generation world model Genie 4, capable of simulating physical processes such as aerodynamics and collisions.

    Google releases next-generation world model Genie 4, which can natively generate music, ambient sound, and interactive dialogue while generating worlds. Compared with the previous generation Genie 3, the new model adds native audio, real-time character dialogue, and multi-step reasoning and scene control capabilities provided by the Gemini model. It is currently being gradually opened to trusted testers in the form of a research preview.

    Why it matters: Genie 4 builds on the previous generation by adding native audio, real-time dialogue, and multi-step reasoning, allowing creators to evaluate its interactive simulation capabilities accordingly.

  3. ComfyUI Wiki · 工作流与开源创作分享63

    Krea 2 Turbo 3-step distilled LoRA released.

    lvladikov released Krea2-Turbo-Distill-3step-LoRA, compressing Krea 2 Turbo from 8 steps to 3 steps, about 2.6x speedup (from 65.4 seconds to 24.9 seconds), with fine texture reaching 1.11x that of the 8-step teacher. This rank-64 LoRA can be loaded directly in ComfyUI, diffusers, and MLX, without custom nodes, and comes with a pure stock node workflow.

  4. Reddit comfyui · 创作与工作流分享68

    PrunaAI releases a 6-step LoRA adapter for Qwen-Image-2.1, without requiring an upsampling step.

    PrunaAI releases a 6-step LoRA adapter for Qwen-Image-2.1, compressing image generation from 40 steps to 6 steps. The 6-steps-upsampled version removes the separate prompt upsampling step, saving 41 to 80 seconds depending on prompt style. The LoRA is loaded directly onto Qwen-Image-2.1, and the pipeline remains unchanged.

    Why it matters: The new adapter compresses image generation from 40 steps to 6 steps and removes the upsampling step, allowing creators to assess whether to replace their existing generation workflow.

Oct 26Mon
  1. 可灵 Kling71

    Kling 4.0 launches on Runware, supporting 30-second 1080p native audio

    Kling officially reposted Runware's message, stating that Kling 4.0 is now live on Runware, supporting videos up to 30 seconds at 1080p with native audio. This version supports multiple keyframes for the first frame, intermediate frames, and last frame, with up to 15 inputs per request, covering keyframes, reference images, reference videos, and elements, and can edit existing clips based on prompts. Pricing starts at $0.133 per second of output, with a 15% discount off the regular price for the next 7 days; the model ID is klingai:[email protected].

    Why it matters: Kling 4.0 launches via the Runware API, offering parameters such as 30-second 1080p with native audio and multi-keyframe input, which can be compared against existing video generation pipelines.

  2. 可灵官方博客67

    Kling Releases Kling 4.0, Supporting 3–30 Second Generation and Omni Reference

    Kling officially released Kling 4.0 on October 15, 2026, supporting single generations of 3–30 seconds, more stable large-scale motion and complex camera movements, and up to 10 keyframe inputs. Omni Reference can combine up to 15 reference assets in a single generation, including 10 images, 5 videos totaling no more than 30 seconds, and 7 elements, and prompts can be used to make targeted modifications to existing assets' subject, background, expression, camera position, and style.

    Why it matters: The official post reviews Kling 4.0's reference asset, keyframe, and targeted editing capabilities, and offers workflow suggestions from storyboard to delivery.

  3. IT之家38

    Musk's Grok releases the first 7 minutes of an AI version of 《The Odyssey》, criticized for lacking artistic appeal

    Musk's Grok, together with Wonder Studios and director Dave Walker, released the first seven minutes of an AI version of 《The Odyssey》, calling it the first episode of the series. TechCrunch commented that its visuals are better than Musk's other recent AI videos, with some sea-voyage segments approaching live-action footage, but that overall it lacks artistry and appeal, with lengthy narration and stiff dialogue that make it hard for viewers to empathize.

  4. Aitrepreneur · YouTube 创作视频58

    Aitrepreneur tests the open-source model MOSS Image and releases a ComfyUI workflow

    Aitrepreneur tested the open-source 5B-parameter image model MOSS Image, saying it can run locally in ComfyUI, generate photorealistic photos and fantasy illustrations, and outperform Krea 2 in several respects. The author also provides a companion workflow supporting standard and Turbo generation, prompt enhancement, and LoRA experiments.

Oct 25Sun
  1. SuSu_酥酥64

    Block releases full-duplex speech model Dessa, runs locally in real time on an M4 Max Mac

    Jack Dorsey's Block released the full-duplex speech model Dessa, which the company calls its most realistic full-duplex speech model to date, with conversations carrying emotion, pauses, and natural rhythm. The model was trained from scratch using only human conversation recordings, without text or pretrained models, can run in real time on an M4 Max Mac or higher, and was trained on a single NVIDIA GB200 NVL72. It is currently open only to security researchers in a closed beta, with a waitlist for public access.

  2. 我爱计算机视觉 · 公开公众号 RSS36

    Open-o3-Video-v2 released: a video agent that thinks while watching and actively retrieves spatiotemporal evidence

    A team from Peking University, Fudan University, and Nanyang Technological University proposed Open-o3-Video-v2 based on Open-o3-Video, turning video reasoning into a multi-round agent process of "discover information gaps → actively crop clips → locate spatiotemporal evidence → incorporate into reasoning."

Oct 24Sat
  1. 可灵 Kling65

    Kling 4.0 Launches on Lovart, Supporting 30-Second Segments and 10 Keyframes

    Kling 4.0 is now available on Lovart, generating clips up to 30 seconds in a single pass, with support for 10 keyframes to control pacing and up to 15 character/product/scene reference assets. The official claims include more stable motion, better audio, and larger camera movement ranges, with Lovart's Arti responsible for generating the shot list.

    Why it matters: Kling 4.0 launches on Lovart, generating 30-second clips per single run with support for 10 keyframes and 15 reference assets, allowing evaluation of long-shot control capability.

  2. 可灵 Kling80

    Kling 4.0 launches on fal with an open API

    Kling 4.0 is now publicly available on the fal platform, with the API open to all customers. The model supports native video generation up to 30 seconds with audio, up to 15 image, video, and element references, and precise control with up to 10 keyframes, with improvements in motion, character consistency, and cinematic storytelling.

    Why it matters: Kling 4.0 opens its API through fal, with native 30-second generation with audio and changes in multi-reference and keyframe control available for creators to evaluate integration.

  3. Monster社区 · 中文ComfyUI实测与工作流22

    Comprehensive review of MiniMax H3 Light Cone Edition VBVR: enhanced prompt adherence, spatial relationships, and physical logic

    A review of MiniMax H3 Light Cone Edition VBVR shows enhancements in three areas: prompt adherence, spatial relationships, and physical logic. The review centers on the version's actual generation performance, focusing on verifying the improvements in the above capabilities.

  4. 极客公园55

    ByteDance Volcano Engine releases Seedance 2.5 4K video generation model

    Volcano Engine officially released Seedance 2.5 4K, officially calling it the strongest model in the Seedance 2.5 generation, enhancing character expressions, subtle movements, and visual details, and supporting 10-bit color depth. This version supports Draft mode, allowing users to first verify storyboards at 480P and then generate 4K final footage; currently it is invitation-only, prioritizing 20 global film and television studios and directors for pre-production of TV series and film shots.

  5. Reddit StableDiffusion · 创作者与工作流分享38

    Moss can natively generate comics and image sequences of the same character in any style

    After merging stripe fix lora and Qwen 9b int8 into Moss SFT int8, it can generate story sequences or comics within a single image while maintaining the same character and supporting any style. The creator says its prompt following is extremely detailed, and with the experimental turbo method using dmd2 turbo lora it is still quite fast, with all images natively 1080p.

  6. SuSu_酥酥78

    Tencent Hunyuan open-sources world model Hy-Nomas, supporting interactive video generation from text and camera trajectories

    Tencent Hunyuan has open-sourced its world model Hy-Nomas, which supports free interaction using text prompts and camera trajectories and has long-horizon generation capability with world memory. The model provides lightweight LoRA adaptation for vertical domains such as robotics; after 8-step distillation it can generate 720p@24FPS video on a single RTX 4090. The weights are now available on Hugging Face and GitHub, along with an online Demo.

    Why it matters: Tencent Hunyuan open-sources world model Hy-Nomas, supporting text and camera trajectory control; 720p video generation runs on a single RTX 4090.

Oct 23Fri
  1. SuSu_酥酥58

    ByteDance Seedance 2.5 4K First Batch Targeted at 20 Top Directors and Studios

    A leak claims that ByteDance's video large model Seedance 2.5 4K has been launched by Volcano Engine for a first batch of enterprise trials, supplied only to 20 leading directors and studios on a targeted basis, without broad distribution. This version strengthens character micro-expressions, performance subtlety and reference consistency; in the workflow, a 480P sample is first used for validation before producing the 4K final cut. Currently 20 works are using it, and as early as next month a satellite TV S-tier IP and a global major IP film will air.