Skip to content
Date and source

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

Nov 2Mon
  1. CoAnimator · 可编辑视频与动画工作流41

    Comparison of Gemini CLI and Claude Code in video production

    Claude Code has documented integration in CoAnimator, while Gemini CLI is more suitable for teams that already have enterprise or paid API access. On June 18, 2026, Google migrated free individual users and Google AI Pro/Ultra subscribers from Gemini CLI to Antigravity CLI.

  2. Stuubzzz · 视频工作流实测75

    MiniMax H3 Character LoRA Training Test: 29 Clips, 4000 Steps, Why Step 3000 Was Chosen

    The author trained a MiniMax H3 character LoRA using AI Toolkit on an RTX 5090, with 29 3-second clips, rank 16, 4000 steps, taking about 15 hours of GPU time, and ultimately chose the step 3000 checkpoint.

    Why it matters: The author tested training a MiniMax H3 character LoRA using AI Toolkit on an RTX 5090, and provides the specific settings of 29 clips and 4000 steps, along with 5 reusable rules for reference when training other video models.

  3. Stuubzzz · 视频工作流实测70

    MiniMax H3 RefMod and reference image testing: long-video facial consistency comparison

    The author's testing found that MiniMax H3's RefMod can improve long-video facial matching from 0.23 to 0.80; the 0.5 MP version scored 0.75 in the second segment compared to 0.25 without RefMod.

    Why it matters: Testing shows that MiniMax H3's RefMod can improve long-video facial matching from 0.23 to 0.80; the 0.5 MP version offers better value.

Nov 1Sun
  1. Reddit StableDiffusion · 创作者与工作流分享26

    Krea 2 Fine-tuning Experiment: Surprisingly Effective, but Sacrificed Versatility

    The author performed seven consecutive fine-tuning runs on Krea 2, training on about 10,000 self-taken photos, resulting in a highly realistic photographic checkpoint, but at the cost of other styles and LoRA compatibility. The training set size was about 300 to 2,500 images, with up to 4,000 steps per run, and at checkpoint 12500 the photographic effect was excellent, but the model exhibited catastrophic forgetting. The author shares this experiment as a reminder that continuous full fine-tuning can lead to over-specialization of the model.

  2. Reddit comfyui · 创作与工作流分享53

    MiniMax H3 Singularity V2 New Workflow: Faster Motion, Better Faces, and Low VRAM Optimization

    Creator cgpixel23 shared an optimized ComfyUI workflow for MiniMax H3 Singularity V2, using INT8 quantization, Turbo LoRA, and optimized attention nodes to reduce VRAM requirements. The model is about 32GB. The new version has improvements over the previous version in fast motion, motion handling, distant faces, close-up portraits, visual style, prompt adherence, and audio reference processing. The author has publicly released the workflow and video tutorial.

  3. note · cocoredo_2509 · 图像视频创作分享56

    Reproducibility of AI production workflows: ComfyUI and multi-model collaboration

    The reproducibility of ComfyUI multi-model workflows hinges on synchronized management of production condition records and quality confirmation. Specific methods include evaluating characters and backgrounds separately, decomposing actions before generation, and comparing models under unified conditions. Costs should include labor and regeneration expenses, and it is recommended to start with small workflows and expand gradually.

  4. 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.

  5. Reddit StableDiffusion · 创作者与工作流分享63

    Krea 2 LoKr Training Test: Testing clothing and character consistency across 10 concert eras at 1536px, and sharing LoKr vs LoRA comparison, training steps, and trigger word tips

    The author trained a 1536px Krea 2 LoKr to test clothing and character consistency across 10 concert eras, and after comparing with LoRA found that LoKr performed better. The optimal training steps were 1,500 steps (37.5 epochs), and using a real name as the trigger word worked better than a synthetic trigger word. The author also shared the ComfyUI workflow, LoKr weights, and 40 showcase images.

  6. Reddit comfyui · 创作与工作流分享45

    Complete music video with AI audio-visual and vocals overlaid, made with LTX 2.5 and ComfyUI

    The author used LTX 2.5 and ComfyUI to create a rap music video about GPUs, with all 52 clips generated locally. 13 Gemini still images each generated 4 LTX 2.5 image-to-video clips, batch-processed on a 5090 at 832x480 resolution, with each 4-second clip taking about 25 seconds. Beat-synced editing aligned the cuts to the drum hits, with hard cuts at vocal entries; the whole process took about 4 hours.

  7. Reddit comfyui · 创作与工作流分享70

    The author used MiniMax H3 and the self-developed Video Builder in ComfyUI to create a Halloween music video.

    The author used the self-developed Video Builder to have Claude automatically complete lyrics, prompts, and the render queue in ComfyUI, producing a MiniMax H3 Halloween music video. The entire video has 19 scenes, 2:29, with 80 renders in total and about 12 GPU hours; the author manually handled story ideas, the Suno song, and Nano Banana scenes, while Claude automatically generated characters, planned shots, and checked lip sync.

    Why it matters: The author published the complete division of labor and 2-pass workflow for Claude-driven ComfyUI music video production, which can be reused in one's own video generation pipeline.

Oct 31Sat
  1. FireCut · 口播剪辑与字幕教程17

    FireCut: 7 Quick Video Editing Tips Every Creator Should Know

    FireCut shares 7 tips for speeding up video editing, covering Premiere Pro's Q/W ripple trimming, Pancake editing, proxy editing, and adjustment layers, among others. In particular, FireCut uses AI in Premiere Pro and DaVinci Resolve to automatically remove silence and duplicate shots, generate subtitles and dynamic push-ins, and add B-roll and background music with one click (including the Storyblocks library).

  2. note · fordtaurus116 · 图像视频创作分享15

    SkyProduction: AI makes 3D spaces and camera movements freely available, using GPT Image 2.5 × Seedance 2.5 to actually produce videos

    SkyProduction combines models such as GPT Image 2.5 and Seedance 2.5 to provide a canvas, 3D space creation, and a camera movement library, allowing users to freely combine features to make videos. For 3D creation, you only need to paste in a background image and it automatically generates a 3D space; camera position and angle can be freely adjusted, or you can directly choose from the preset camera movement library. The experience is like being a film director, and the production process is free and fun.

  3. note · cocoredo_2509 · 图像视频创作分享68

    Pika H3 Flex: Import, Quality, and Cost Verification for Video Creators

    The author verifies Pika H3 Flex from a production-site perspective, pointing out that official performance cannot be directly equated with actual usability; character consistency, audio track separation, and 4K upscaling need to be evaluated separately. H3 Flex is based on MiniMax H3, supports up to 12 reference materials and 5 types of audio track separation, and after generating at 480p/768p can be upscaled to 1080p/2K/4K, but 4K is upscaled rather than natively generated.

    Why it matters: On-site verification of Pika H3 Flex shows that character consistency, audio track separation, and 4K upscaling need to be evaluated separately; the author's three short tests and recording method can be directly used for other model adoption decisions.

  4. Reddit comfyui · 创作与工作流分享58

    Creating a short animation scene from a character design sheet using FastH3 and ComfyUI

    The author used FastH3 and ComfyUI to create a ten-second animation scene from a character design sheet, and published the workflow, prompts, and reference assets. The workflow uses a custom Krea 2 Wulver fine-tune to generate the character design sheet, and FastH3 OmniRef generates a ten-second square output with eight-step sampling. Mochi was left out entirely. Custom model weights are not included, but the provided reference PNG can be used starting from the animation stage.

  5. Curious Refuge 创作教程69

    How to Use AI to Relight a Video (Step-by-Step Tutorial)

    Curious Refuge published a tutorial demonstrating how to use AI to relight videos, and tested the workflow's effectiveness on live-action footage and gray-box studio scenes. The workflow first exports still frames and relights them using an image model (Nano Banana 2.1 or GPT Image 2.5), then uses the relit images as references to apply to the original video. Testing found limitations such as exposure shifts, banding, and resolution loss, with output resolution lower than the 4K source.

    Why it matters: This workflow uses static-image relighting as a reference for the video model, allowing creators to evaluate the boundaries and limitations of AI relighting in post-production.

  6. Higgsfield 官方博客76

    Higgsfield's real-world test of 2026 AI video costs: a 10-second finished video ranges from $1.35 to $7.85

    Higgsfield's test shows that a single generation of a 10-second AI video costs about $7.85 (160.125 credits) on Seedance 2.5 at 1080p, and about $1.35 on Kling 4.0 at 720p. The actual cost must also account for retakes, image preparation, and post-production.

    Why it matters: Provides a method and budget formula for calculating the true cost of AI videos under per-generation pricing, allowing creators to plan retake counts and monthly output accordingly.

  7. Reddit comfyui · 创作与工作流分享57

    MOSS 2.5D Anime generates an image in about 62 seconds at 1024x1536 on 8GB VRAM.

    The author tested running MOSS 2.5D Anime on a laptop with 8GB VRAM, generating an image at 1024x1536 resolution in about 62 seconds. The text encoder occupies 9.75GB and needs to be placed in system memory, requiring at least 32GB RAM in practice; the model is licensed under Apache 2.0, outputs can be used commercially, and no custom nodes are needed—only core ComfyUI.

  8. Podcastle37

    How to convert AI-generated 3D models into social media videos using Meshy and Async

    The combination of Meshy and Async can convert AI-generated 3D models into social media videos. The process is: generate a 3D model in Meshy and render it as a 2D image, then use Async to animate the image, add music and subtitles, and output to Instagram, TikTok, or YouTube Shorts. For short videos, Meshy and Async are sufficient; when fine animation is needed, Blender or DeepMotion can still be considered.

Oct 30Fri
  1. Reddit StableDiffusion · 创作者与工作流分享67

    PxTicks open-sources Minimax H3 motion control workflow and companion applications

    PxTicks has open-sourced the Minimax H3 motion control workflow, which allows controlling motion by dragging objects, and comes with background separation, motion path, and mask generation tools that can be directly used with the ComfyUI interface.

    Why it matters: The author has open-sourced the Minimax H3 motion control workflow, which allows controlling motion by dragging objects; creators can refer to it to achieve similar effects.

  2. Reddit comfyui · 创作与工作流分享59

    AI animated comedy QUiBBLE after three episodes: character consistency remains the biggest challenge

    In the third episode WANTED of AI animated comedy QUiBBLE, the author used MiniMax H3 and LTX through ComfyUI, and found that character consistency remains the biggest challenge. The author observed that H3 performs better in restrained acting and dialogue consistency, while LTX is more prone to exaggerated expressions; clothing details and environmental continuity are also difficult to maintain, and they are turning to limited animation-style methods such as fixed camera positions, repeated backgrounds, and character references.

  3. note · nobinlog · 图像视频创作分享50

    Anima Reference Adapter for ComfyUI Test: To What Extent Can Uploaded Assets Change Things?

    The Anima Reference Adapter for ComfyUI test shows that the tool can change pose, background, and camera angle via prompts while preserving the reference image's character face shape, hairstyle, eyes, and clothing. The art style can only be partially inherited. When adding character prompts such as Hinata Hyuga, feature fusion occurs. Its main uses are reproducing original characters, inheriting unique clothing, and continuously generating the same character.

  4. MindStudio Blog · 创作分享70

    3D to AI Video Workflow: How to Maintain AI Character Consistency

    The 3D-to-AI video workflow uses 3D models as references to solve the problem of AI video characters drifting as camera angles change, involving model generation tools such as Trippo 3D, Hunyuan 3D, and Meshi V7. Among them, Meshi V7 has flexible export formats, Cozy Clay allows manipulating models in the browser, and Simulon can place characters into real environments to generate videos. However, the generated models are not rigged with skeletons by default, and animation still requires additional rigging steps.

    Why it matters: A workflow using 3D reference models to solve AI video character drift, and comparing the applicable scenarios of tools such as Trippo 3D, Hunyuan 3D, and Meshi V7.

  5. Reddit comfyui · 创作与工作流分享66

    MiniMax H3 Fight Test: ComfyUI Workflow with First Frame plus 18-Frame Storyboard and 3 LoRAs | RTX 3060 12GB

    The author tested MiniMax H3 in ComfyUI, using a first frame plus an 18-frame storyboard and 3 LoRAs to generate a continuous fight sequence, outputting 1536×864 at 24fps for about 12 seconds (289 frames) on an RTX 3060 12GB, with generation taking about 28 minutes. The result was a continuous action sequence with no slow motion or cuts, but the movements still looked unrealistic, especially ground interactions; the author did not test each component separately, so this is not a comparative conclusion.

    Why it matters: The author published the complete parameters for generating a continuous fight sequence on an RTX 3060 12GB using a first frame, an 18-frame storyboard, and 3 LoRAs, which can be transferred to similar action scenes.

  6. 可灵官方博客29

    What is an Animatic and how to create one with Kling AI?

    An animatic is a moving preview created by arranging storyboard frames along a timeline, used to assess pacing and editing points before production. Kling AI's Kling 4.0 video generation and editing model supports video generation from 3–30 seconds, and with Omni Reference it can combine up to 15 reference materials (10 images, 5 clips totaling 30 seconds of video, 7 subject references), helping creators quickly generate dynamic drafts from storyboards and adjust action, camera movement, and scene pacing.

  7. Leeveo AI Tutorials · 创作教程与项目分享65

    How to create consistent 3D cartoon characters with ChatGPT Images

    Lee Veo published a tutorial on using ChatGPT Images and structured prompts to create consistent 3D cartoon characters without any 3D modeling experience. The tutorial covers the complete workflow for character sheets, expressions, poses, scenes, and image-to-video animation, and provides reusable prompts.

    Why it matters: This workflow of using ChatGPT Images to generate character sheets and maintain consistency across images can be directly applied to AI short dramas or series animation production.

  8. 影视独舌 · 搜狐公开专栏58

    2xLabs won a million-yuan prize from Douyin for 'Helong' and was recognized by the Film Bureau. Will they become the next Xin Shuang?

    2xLabs won the gold award at the Douyin AI Creation Competition with its AI short film 'Helong' and was invited to participate in the 9th China Film New Forces Forum hosted by the Film Bureau of the Publicity Department of the Central Committee of the Communist Party of China. The team has years of experience in traditional film and television production. The AI imagery for the 17-minute short film took only about a week to produce, with computing costs under 100,000 yuan. Twelve people participated in the creation. The soundtrack is original music, and all voice acting was generated by AI. The team's goal is to become a leading AI film and television content production institution, undertaking commercial projects, producing external IPs, and developing original series.

  9. 王若风 · 开源创作项目分享75

    Without touching video generation models, Lemo-Opuscar has Claude write code to finish 59 short films.

    Lemo-Opuscar is an open-source project that uses Claude Code to write code to generate short films, completely independent of video generation models. It renders frame by frame via canvas/WebGL, screenshots with headless Chrome, and composes with ffmpeg. It has accumulated 59 styles and received 8,274 stars in one month.

    Why it matters: The author distilled a three-step blind director mode from Lemo-Opuscar: turn the work into pure functions, swap senses for assertions, and write taste into documentation. This can be transferred to other AI creation tasks.

  10. Reddit comfyui · 创作与工作流分享48

    ComfyUI Animated Short Film: Sharing My Creation

    A creator used ComfyUI to produce the first episode of a 3-episode animated mini-series, telling the story of a little witch being consumed by a curse and meeting a giant cat. The technical workflow includes using ChatGPT to generate poses and compositions, a Qwen Image workflow to remove the ChatGPT style, MinimaxH3 to complete the animation, and ElevenLabs to handle music and sound effects. The author is seeking feedback on character consistency and multi-shot animation.

  11. note · ktcrs1107 · 图像视频创作分享23

    HIX AI Creates Halloween Video! Hero Wu

    HIX AI's Seedance 2.5 model was used to create an approximately 18-second vertical Halloween short drama titled "Trick or Treat, Here's My Business Card." The protagonist, otherworldly hero Wu, hands out business cards instead of candy to children on Halloween night. During production, it was discovered that trimming the story was more effective than adding prompts for enhancing the short drama's impact, and ultimately only the core scene of handing out business cards on Halloween night was retained. As a platform integrating multiple AI functions, HIX AI also offers image generation, AI chat, and slide presentation creation features.

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

    A clever trick for generating longer AI videos: generate backwards from the end

    Reddit user MustBeSomethingThere shared a technique for generating longer AI videos, aiming to reduce quality drift in segmented generation of long videos by first generating the beginning, then using the beginning as a reference to generate the end, and then generating backwards from the end toward the middle. For a six-segment video, the generation order is 1→6→5→4→3→2, but the playback order remains 1→2→3→4→5→6.