Skip to content
Date and source

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

Oct 14Wed
  1. Reddit StableDiffusion · 创作者与工作流分享22

    Tip: extract the original ComfyUI workflow from an old image

    After generating images with Krea2 in ComfyUI, a user used Claude to build a prompt and workflow extractor: drag an image in to export a .json, then drag it back into ComfyUI to restore the complete workflow used at the time. The user said no programming background is needed—just describe the requirements and iterate.

  2. note · dreammii · 图像视频创作分享22

    What to do when AI video BGM is too loud? A "three-layer sound" organizing technique even beginners can use

    To address the problems of unclear narration, abrupt sound effects, and overly loud BGM in AI videos, you can use the "three-layer sound" organizing method: first set the lead voice (narration or AI music), let BGM and ambient sound serve as support, and use sound effects only at transitions.

  3. note · dreammii · 图像视频创作分享22

    All AI images look the same? Use "4 design images" to differentiate note thumbnails and illustrations

    To address the problem that when AI generates images for note articles, the thumbnail and 3 illustrations easily look the same, you can use "4 design images" to first assign each image a different role and scene: the thumbnail conveys the theme, illustration 1 covers the overall composition, illustration 2 covers the method of operation, and illustration 3 covers completion confirmation. When generating, split the prompt into common conditions (such as a cute Japanese adult woman, bright anime style, blue-pink tones, about 5,500K neutral daylight, 16:9) and individual scenes for each image, and generate and check them one by one; illustrations should explicitly not include large headline text.

  4. CoAnimator · 可编辑视频与动画工作流30

    Can OpenClaw make videos? From chat brief to finished MP4

    OpenClaw can coordinate video tools to produce video: it can fetch generated clips from a configured video generation provider, and it can have the agent build and render an editable project in CoAnimator. Telegram handles sending instructions and receiving previews, while actual editing and rendering happen on the computer connected to the project; before export, the storyboard plan, still frames, and rough cut must be confirmed in order. This integration flow has not yet been tested end to end.

  5. ArchiGen AI · 图像生成与ComfyUI实战30

    How to set a "change budget" for ComfyUI render enhancement

    ArchiGen AI proposes using a "change budget" to guide ComfyUI architectural render enhancement: before building the workflow, list the protected architectural facts, the permitted image changes, and the edits that require masking. Global passes suit overall atmosphere adjustments like exposure and color and need low denoising; changes with clear ownership, such as materials and reflections, should go through masks or local inpaint, and every output should be compared against the original image.

  6. ArchiGen AI · 图像生成与ComfyUI实战20

    How to make geometric control in AI renderers measurable

    ArchiGen proposes replacing qualitative labels like "medium" with verifiable architectural invariants to evaluate geometric control in AI renderers: before generation, select 8-12 visible invariants, run at least 6 fixed-seed outputs from the same source image, score each as preserved, uncertain, or failed, then weight failures by project stage and record fix time. Geometric control is the acceptance pass rate under this test, not a weak/medium/strong label.

  7. note · 7god_ebisu8g2sdr · 图像视频创作分享26

    A PC beginner's 5 failed attempts at fruit-cutting ASMR animations with ChatGPT and AI video generation

    A PC beginner used ChatGPT and AI video generation to make fruit-cutting ASMR short videos and documented 5 failures: a wrapped knife passing through grapes without cutting them, cut flesh not matching the pre-cut appearance, a burger video with two knives and a cutting direction misaligned with the cross-section, a glass banana's cross-section appearing suddenly, and a female character added to the scene along with an extra small woman and coffee cup. The author concludes that AI video struggles to naturally connect object motion with before-and-after states, and that the pre-cut state, cutting action, and post-cut cross-section must be specified explicitly, with repeated generation and prompt fixes.

  8. Wireflow · AI短剧与视频制作教程15

    Design.com vs. Canva vs. Adobe Firefly: which AI image generator is better for beginners?

    Testing with the same set of three prompts showed Design.com is the best choice for beginners, Canva suits existing Canva users, and Adobe Firefly is better for users in the Adobe ecosystem. Design.com supports everyday-language prompts, AI editing, and prompt optimization, and connects to 50+ design and AI tools to complete branded deliverables; it can generate free without an account, with paid plans from $3/month. All three tools produced usable images with simple, detailed, and marketing prompts.

  9. Hugging Face Daily Papers25

    StreamErase: toward ultra-fast streaming video erasure

    StreamErase achieves streaming video erasure with sub-second latency and up to 26 FPS on a single GPU, enabling real-time interactive removal of target objects from video. The method first progressively transfers the erasure capability of a bidirectional base model to an efficient unidirectional streaming eraser, and employs sliding-window causal attention with attention sinks and KV cache rolling to achieve constant-speed generation under a fixed context window.

  10. Hugging Face Daily Papers29

    WorldCast: a multi-step world state prediction benchmark, GPT-Image-2.5 scores 72.8

    Researchers propose the WorldCast benchmark, containing 320 hand-designed cases across 8 categories, to evaluate multi-step world state prediction tasks, testing 12 representative models in total. GPT-Image-2.5 System scores 72.8, showing preliminary real-world visual state prediction ability; Nano Banana Pro scores only 42.8, and the best open-source model scores 20.1. The experiments also find that strong generative models can effectively leverage the capabilities of their underlying language models.

  11. Hugging Face Daily Papers22

    RepForcing: representation-space initialization for autoregressive video distillation

    RepForcing proposes replacing intermediate teacher trajectory regression with representation-space supervision to learn causal initialization for few-step autoregressive video generators, reusing synthetic videos generated by bidirectional models without extra causal ODE trajectory generation. Under four-step sampling, after DMD the method raises VBench Total from 84.04 to 84.66 (+0.62) and VisionReward from 6.326 to 10.074 (+59.25%).

  12. Hugging Face Daily Papers27

    Orbia: can generated video form a consistent 3D world?

    Researchers introduce Orbia, the first benchmark to systematically evaluate the 3D consistency of interactive world models, assessing along three dimensions: input preservation, persistence of generated content, and 3D self-consistency. The benchmark brings together public datasets, self-collected scenes, and simulations, building 720 indoor and outdoor instances and six exploration-revisit protocols, covering 27 text-, action-, and camera-conditioned models.

  13. Hugging Face Daily Papers22

    Self-conditioning enables deep sequential reasoning in Flow Matching

    The study finds that the trajectories of a native flow matching sampler can be simulated by a constant number of parallel model evaluations, and that increasing sampling steps cannot improve the expressive power of a fixed-depth backbone model. Self-conditioning (using the previous step's estimate as a condition) breaks this bottleneck: across multiple visual reasoning tasks, flow matching with self-conditioning achieves higher accuracy and greater reasoning depth, and performance improves as sampling steps increase; theoretically, a fixed-depth backbone can reach the expressive power of a space-bounded Turing machine.

  14. Hugging Face Daily Papers22

    Speculative Sampling in interactive video generation

    Addressing the reliance of interactive video generation on small bidirectional chunks to reduce latency, researchers propose Speculative Sampling, which extends the modeling window by using noise latents as padding and caches predictions for reuse when conditions are unchanged, thereby combining the image quality of a large bidirectional window with the low interaction latency of small chunks. Latency analysis shows the method eliminates the lower bound on waiting time caused by chunk boundaries, making the bidirectional window size a decision variable that can jointly optimize image quality and sampling time, and enabling large bidirectional models to be served in interactive streaming systems without complex streaming post-training.

Oct 13Tue
  1. TelosCompass · note AI视频创作实践36

    How to use prompts to set camera movement in AI video generation and lock in the final frame

    When generating AI video from a single photo, camera movement only requires choosing one of three options—static, dolly in, or dolly out—and specifying the stopping position to determine what the final frame will show. If you only write "dolly in," the frame may stop on a close-up of a vase pattern without revealing it's a vase; dolly out may cause the AI to generate sheds and windows not present in the photo, so real-world scenes should use a wide-angle photo with "static." DomoAI's official blog recommends using only one camera movement per scene, while Runway's official documentation advises adding speed or duration to camera movement terms and distinguishing between camera push-in and zoom.

  2. note · singtothesky79 · 图像视频创作分享8

    Starting AI video creation on my own ②: Using Gemini to generate a novel-length script and turn it into a video

    A self-taught AI video creator tried inputting a long text into Gemini to generate a story and then produce a video, asking Gemini to write a script roughly the length of a novel chapter, telling the story of two adventurers who, after a thirty-year journey, ultimately fight each other. After generating the video, the creator said the result differed from expectations.

  3. note · aiholic0831 · 图像视频创作分享22

    ChatGPT generates a "20-something female engineer" character sheet: 7 prompt sets lock in facial features and body type

    The author first used ChatGPT image generation to create a character sheet, then included it as a reference in each of 7 scene prompts to lock in the same 20-something Japanese female engineer's face shape, hairstyle, glasses, and body type. Scenes include coding, online outage, online meeting, employee badge selfie, foggy glasses, Akihabara holiday, and death march. All images are unified in 3:4 vertical format, and the character sheet can be saved for direct reuse.

  4. note · aicu · 图像视频创作分享36

    AICU Magazine Vol.29 special feature on GPT-Image-2.5: character reproduction, Flare/Sunburst division of labor, and API practices

    AICU Magazine Vol.29 is out, with a cover feature focusing on the new image generation model GPT-Image-2.5, using the official character AiCuty's AiCutyBench to test its character reproduction "yield rate." The issue also covers the use-case division between the high-speed model Flare and the high-precision model Sunburst, translations of the GPT-Image-2.5 API manual and prompt guide, as well as Suno v6 partial editing and UMG's multi-year licensing agreement with ElevenLabs.

  5. note · natty_mimosa8652 · 图像视频创作分享30

    How to Create Original AI Beauties: A Practical Guide to Escaping Mass-Produced Faces and Generating Unique Worldviews and Personalities

    A practical tutorial explains how to use tools like Stable Diffusion and Midjourney to generate original AI beauties that aren't "mass-produced." The core method involves first designing a "character and worldview" design sheet that includes age, bone structure, personality, charm points, and visual symbols, then writing prompts in a five-layer structure. The tutorial recommends removing words like "extremely beautiful" and "cute girl" from prompts, instead using specific descriptions of features, lighting, composition, and texture to convey beauty, and provides ready-to-use character design templates.

  6. note · noble_mink9287 · 图像视频创作分享22

    College Student AI Side Hustle #3: Four Tools for Beginners to Make Videos with AI (PixVerse, CapCut, TikTok, etc.)

    A college student created a "Tokyo 100 Years from Now" themed video by using AI image generation tools for assets, then converting images into dynamic footage with PixVerse, and finally editing and assembling with BGM in CapCut before publishing on TikTok. The author notes that PixVerse often produces erratic camera movements and distorted buildings, requiring fixed camera positions and scene-by-scene instruction adjustments; they also emphasize that being able to make videos and being able to earn a side income from them are two different things.

  7. note · rosy_dog6649 · 图像视频创作分享24

    What to Do When Japanese Text in AI Images Breaks Down? Add Text in Post-Production with Adobe Express to Improve Readability

    To address issues with incorrect Japanese kanji and awkward line breaks in AI-generated images, use a "generate a text-free background first, then add text with Adobe Express" workflow: during generation, leave bright blank spaces at designated positions and avoid rendering text, then add editable text boxes in the Text panel, setting the title and date separately. When adjusting, unify font size hierarchy, increase contrast between text and background, and break lines by meaning. Before exporting, check for typos and numbers at full size, and confirm the title is readable at a glance at small sizes.

  8. note · psych_ai_it · 图像视频创作分享22

    Recreating Nursing Scenes with AI Video: The Harder Part Than the Visuals Is "Sound"

    An IT engineer with a psychology background used Grok's video generation feature and AI tools like Google Flow to recreate Japanese nursing care scenes, finding that the visuals are already quite natural but sound has become the biggest challenge. Even after specifying Japanese dialogue, the model may still output English, and in loud scenes like "I want to go home!" the AI voice may hit the right volume but struggles to convey emotions like anxiety and confusion, sounding exaggerated or mechanical. The author believes that the more realistic the visuals, the greater the responsibility for sound expression, and that in the future, video and audio may need to be generated separately.

  9. Frank Houbre · AI电影制作实战22

    AI Video Tenders: How Freelancers Can Bid

    Under France's 2026 public procurement rules, AI video freelancers can bid on video production tenders issued by local governments, tourism boards, and others. Service contracts below €60,000 do not require open tenders, those between €60,000 and €90,000 use an adapted procedure, and those above €90,000 must be published on BOAMP or in official legal gazettes. BOAMP's free keyword search can combine "vidéo" with "Services" and "Avis en cours" for filtering, and video is often just one of multiple lots in a tender.

  10. 有机大橘子34

    AI Beginner's Guide: Making AI Story Short Films All-in-One with Miora Weaving Agent

    A blogger used Miora Weaving Agent to turn the "Qian Hong Mao" (Red Cat) story from the Southern Song Dynasty's "Yijian Zhi" into a 1-minute-18-second short film, generating 50 video segments and spending approximately 4,000 credits (about $40). The Agent broke the story down into framework, outline, script, character cards, storyboards, and shot assets, using the Seedance Fast model for the video portions. Red cat consistency remains the main issue, requiring character cards and repeated generation attempts to adjust, and it's still far from one-sentence story short film generation.

  11. note · rouageailab · 图像视频创作分享62

    Hands-On Test of MiniMax H3's Character Swap LoRA: Replacing Live-Action Dance with Animated and Claymation Characters

    The author tested the Character Swap LoRA v1.0 for MiniMax H3 released by Alissonerdx on October 12, combining 3 Pexels videos with 3 characters (themselves, a 2D animated girl, and a claymation figure) to produce 9 results, 8 of which could keep up with the original video's movements. The claymation figure was even able to complete floor support and leg kicks in full-body dance sequences.

  12. note · sozaixx · 图像视频创作分享60

    Hands-On Test of ComfyUI on an 8GB VRAM Laptop: Generate Large Images First Then Upscale, and Use EasyCache to Cut Video Generation from 37 Minutes to 21 Minutes

    The author tested VRAM and caching strategies on ComfyUI 0.39 with an RTX 5060 Laptop (8GB VRAM, 32GB RAM). Z-Image Turbo did not run out of memory at 2560×2560, but took 189 to 213 seconds, whereas generating at 1024×1024 first and then upscaling to 2560 via RealESRGAN x4 took only about 12 seconds and produced more stable compositions.

  13. 机器之心 · 公开公众号 RSS36

    Teams from the University of Oxford and others propose the interactive video world model Oneira, which uses a World State Table to record object states

    Teams from the University of Oxford, Monash University, and the University of Bristol propose the interactive video world model Oneira, which uses a World State Table maintained by a Coding Agent to record scene objects and their states, separating 'world state' from 'frame generation' for processing.