Skip to content
Date and source

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

Oct 15Thu
  1. note · fordtaurus116 · 图像视频创作分享8

    ChatGPT draws four characters against an autumn foliage background: expressing charm through gaze and gestures | AI illustration

    This ChatGPT AI illustration share draws 4 characters against an autumn foliage background, with the focus on adjusting face orientation and pupil direction separately—merely fine-tuning the gaze conveys different personalities such as gentle, wistful, or playful. The author says that within the same background, simply changing the gaze and hand gestures lets each character show a different expression.

  2. Hugging Face Daily Papers24

    AVC-Reward: An Event-Centered Audio-Video Consistency Evaluation Framework

    Addressing multi-event combinations, multi-layer audio mixing, and asymmetric cross-modal relationships in joint audio-video generation models, researchers proposed the event-centered evaluation framework AVC-Reward, which decouples audio and video events at the event level and performs cross-modal cross-validation, aggregating into a coarse-to-fine three-level consistency metric.

  3. Hugging Face Daily Papers22

    Dusting the Phantom: allocation strategies for decaying content in dynamic Gaussian representations

    To address the problems of fast-moving small objects being easily lost and details in highly dynamic regions being hard to recover in dynamic Gaussian representations, the research proposes two strategies, Seed Injection and Temporal Biased Relocation, which use temporal appearance changes and reconstruction error to locate under-reconstructed regions, and use time-resolved optimization signals to guide Gaussian relocation, respectively. Without changing the underlying representation and while maintaining a fixed Gaussian budget, the two improve reconstruction quality on multiple dynamic scene datasets.

  4. Hugging Face Daily Papers22

    Mine, Sample, Render: test-time diversity alignment for generative models

    Researchers propose MINE-SAMPLE-RENDER (MSR), a test-time method that mines interpretable categorical patterns from a corpus of real samples and samples them statistically, then uses the generative model only as a conditional renderer to output the sampled configurations, thereby improving generation diversity. On tasks including weather report generation, user simulation, web design, text-to-image, and agent-generated 3D assets, MSR improved alignment with the corpus distribution and diversity of outputs.

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

    Questions About Qwen Image 2.1 LoRA: Can Generation LoRAs Be Used for Editing, and Vice Versa?

    A user asked whether Qwen Image 2.1 LoRAs can work across both generation and editing modes, such as using a generation LoRA to add Doraemon to an image, or using an editing LoRA to generate a person in a specific pose. The user believes Qwen Image 2.1 is a unified model and should theoretically be feasible, but wants to confirm whether anyone has actually tested this.

  6. Reddit StableDiffusion · 创作者与工作流分享18

    How to Reproduce nano banana 2's Generation Results Locally?

    A user tried to reproduce nano banana 2's paid generation results locally using krea 2, flux 2 dev, flux 2 klein 9b, and qwen image edit 2.1, but none came close. Their need is to input two or more face reference images to generate a scene, then convert the result to video using h3 minimax. The video step works fine, but they hope the starting frame can be generated locally instead of paying for it.

  7. Reddit StableDiffusion · 创作者与工作流分享34

    Dissecting Krea-2: What Can You Discover by Adjusting Weights Block by Block?

    A creator used the Arthemy Krea-2 Tuner to adjust Krea-2's model and text encoder weights block by block, trying to find "knobs" that change style without training, but after thousands of renders found only one. Text encoder edits barely damage image quality but have poor consistency — only 4 out of 20 edits made different characters change in the same direction. Commas appear to act as semantic boundaries, with adjacent 2–4 words moving as a group.

  8. Reddit StableDiffusion · 创作者与工作流分享20

    How to Make AI Videos Look Like Casual Phone Footage Instead of Cinematic Shots?

    A user generating "casual desktop product shots" videos with self-hosted MiniMax H3 found that no matter how they prompted, cinematic zooms and camera movements always appeared — even adding smartphone/handheld descriptions and trying with and without realism LoRA still resulted in overly polished output. They asked whether this natural phone footage effect mainly depends on the model, prompts, reference images, or settings, and hope to see a working workflow.

  9. 我爱计算机视觉 · 公开公众号 RSS37

    NeurIPS 2026 | OmniMem enables long video generation to directly retrieve the complete history

    Northeastern University and Adobe Research have proposed OmniMem, achieving full-history long video generation for the first time and allowing the model to retrieve the memories it currently needs in real time. The method keeps full-resolution historical KVs on the CPU, recalls them to the GPU on demand, and customizes Triton and CUDA kernels around block-level retrieval.

Oct 14Wed
  1. note · singtothesky79 · 图像视频创作分享10

    Self-taught AI video creation ③: Using Gemini to make CANVA-generated AI characters move

    After generating AI characters in CANVA, the author tried having Gemini make a video using those characters, and successfully produced an animated clip of multiple characters in different colors playing. The author believes that if this workflow is further refined it could be used to make animation, but progress is currently stuck for the time being, and exploration will continue.

  2. note · kazu_fuwa · 图像视频创作分享76

    MiniMax H3 local deployment hands-on test: comparison of 7 configurations at 16GB VRAM, a 5-second video takes about 2 minutes 40 seconds

    The author ran MiniMax H3 with ComfyUI on a PC with 16GB VRAM and 32GB RAM. After comparing 7 configurations, a 5-second video was completed in as fast as about 2 minutes 40 seconds, while the official 20-step setting takes about 12 and a half minutes.

    Why it matters: The author tested 7 configurations on a machine with 16GB VRAM and provides reusable speed-up combinations and a record of pitfalls encountered.

  3. TelosCompass · note AI视频创作实践34

    How to keep a product's color and shape consistent across two scenes in AI video generation

    When producing AI video generation scene by scene, the prompt written for the previous scene is not automatically carried into the next scene, so the product's color and shape can easily look like a different item after splicing. The approach is to write the unchanging features (such as pale green color, rounded square shape, leaf-shaped engraving) as a "共通メモ" (shared memo) and reuse them verbatim in each scene's prompt, changing only the display range and action. Before generating the video, compare the two scenes' images side by side and check them against the real product, then decide whether to splice.

  4. note · quirky_hippo6791 · 图像视频创作分享15

    First experience with ComfyUI: moving from ChatGPT comic generation to node-based experimentation

    A creator who had been generating comics with ChatGPT for several months installed ComfyUI on a Mac mini and found that it can combine models and generation conditions and compare results item by item, like a "laboratory for image-generation AI." He plans to use it to verify which rules in his accumulated "generation bible" are actually effective, but admits that at present he has only installed the software and has not yet begun researching model selection, node construction, or video generation performance.

  5. FireCut · 口播剪辑与字幕教程22

    FireCut guide: how to automate the video editing workflow

    FireCut released a guide for content creators on automating video editing, stating that AI tools can compress what originally took about three hours of rough-cut cleanup to under three minutes, automatically removing silence, filler words, and identifying inappropriate words. For subtitles, it can automatically generate dynamic animated subtitles, highlight emphasized words, and insert emoji, and B-roll and zooms can also be automatically matched and added based on the transcript text.

  6. note · ya_ma_da_pon · 图像视频创作分享22

    ChatGPT Portrait Generation: 5 Prince-Type "Idol Costume" Prompts for AI Boyfriend Character Art

    The ninth installment of making AI boyfriend idol character art with the ChatGPT image generation model, offering prince-type idol costume prompts, with options such as basic form, long coat, short cut, cape and cloak, and vest. The prompts are used together with the three-point set of "face close-up, full body, three-view," and simply sending them allows free generation of poses and compositions while preserving the character's costume, hairstyle, and facial features; it also supports making unified costumes for idol groups.

  7. Reddit StableDiffusion · 创作者与工作流分享50

    How to Achieve a Transparent Background with Your Own LoRA: Render on White, Repaint on Black

    When making 2D game assets with FLUX.2 klein 4B plus a style LoRA, the author used the old film matting method (Smith & Blinn 1996) of "generate on a white background, then repaint onto a pure black background using the original image as reference" to obtain a transparent background, alpha = 1 - (white - black) / 255, with colors taken from the black-background image to avoid white fringes.

  8. note · tsugiito_design · 图像视频创作分享30

    Making a "Vampire Holding a Pumpkin Lantern" in ComfyUI: Animagine XL 3.1 Fangs, Gothic Costume, and Lighting Tags Plus JSON

    Using ComfyUI with Animagine XL 3.1 to generate a purple-haired, red-eyed vampire holding a pumpkin lantern, specifying one fang on each side of the upper jaw, one hand holding the stem and the other arm supporting the bottom, with two JSON segments attached: one for overall generation and one for local correction of the upper-right wing.

  9. note · kt_kt_89 · 图像视频创作分享24

    After Handing Video Generation to an AI Agent, I Let Codex Take Over the Operator Role

    In addition to a local video generation agent, the author introduced Codex as a higher-level operator, letting it read logs, state, and Reviewer correction instructions, decide on sending back for revision and resuming from checkpoints, while the author only sets policy. This division of labor has been trialed since v0.5.64; finalized execution and safety state management remain local, and boundaries such as budget caps and retry limits are unchanged. The author emphasizes that this is a personal hands-on impression, not a quantified benefit, and that Codex itself also has usage and execution limits.

  10. Reddit StableDiffusion · 创作者与工作流分享36

    Reminder: raising CFG on distilled few-step video models makes output worse, not more faithful to the prompt

    On LTX-2.5, raising CFG from 1 to 4 caused image quality to collapse, with oversaturation and texture distortion. The reason is that the model is a guidance-distilled model, prompt adherence is already baked into the weights, and base inference runs only 3 denoising steps; raising CFG deviates from its operating point during distillation. The author reminds those migrating from the SDXL/SD1.5 habit of CFG 7 that they need a different mindset.

  11. note · aiby_masuhiyo · 图像视频创作分享22

    Generate AI illustration prompts with the "Illustration Parts Gacha" tool: hands-on tests with ChatGPT and Midjourney

    The "Illustration Parts Gacha" tool developed by Akari no Tabibito combines scenes, expressions, poses, clothing and small events into image-generation AI prompts, and supports pinning favorite items. The author generated two sets of prompts with the tool and produced images in ChatGPT and Midjourney respectively; the Midjourney version required fine-tuning the English prompts, while the ChatGPT version was used as-is.

  12. note · kiryu_mizuki · 图像视频创作分享8

    ComfyUI beginner's log: building a workflow from scratch to generate images, with models and prompts the biggest hurdles

    An AI image generation beginner shares the experience of building a workflow from scratch in ComfyUI, using a "cooking in the kitchen" analogy for the relationship between workflows, nodes, models and LoRA. The author picked a model from the SD1.5 model list on Civitai; the first generation produced only a blurry image due to insufficient steps, and after adjustments it took about an hour to produce a fairly decent male image.

  13. Reddit StableDiffusion · 创作者与工作流分享20

    Hands-on test of generating synthetic aircraft images with SDXL + ControlNet: throughput reaches 6.36 images/sec on 4 RTX 5090s, but image quality lags behind

    A synthetic aircraft image generation pipeline was built with SDXL + ControlNet Canny and tested on 1, 2, and 4 RTX 5090s: throughput on 4 cards was 6.36 images/sec, a 3.77x speedup over 1.69 images/sec on a single card, for a scaling efficiency of 94.2%. However, although the aircraft in the generated images are recognizable, the airport ground shows strange rock-like textures, and the author still cannot confidently use them to train an aircraft detector.

  14. note · mintjuku · 图像视频创作分享22

    Kling AI beginner notes: easy-to-understand prompt examples for video generation AI

    A user shared tips on using Kling AI to generate fun educational videos: the simpler the prompt, the better; First Take works best, and repeatedly adding requirements instead causes the visuals to break down. Background, expressions, and camera angles are suitable for specific specification, but when only dialogue is given, the AI may become a narrator rather than the character speaking. Video duration can be set to under 15 seconds; too short will cause overly fast speech or a truncated ending; AI's reading of Chinese characters is incomplete, so it is recommended to write uncertain characters in hiragana.

  15. PixVerse 中文官方博客38

    Kling AI 4.0 prompt guide: camera, motion, and examples

    The Kling AI 4.0 prompt guide proposes a four-part structure: subject, action, scene, camera and style, and emphasizes writing the subject and main action first, with camera and emotion serving the action. The guide recommends writing camera movement instructions with specific terminology, such as "camera slowly pushes in to the subject's face" and "shoulder-height tracking shot," and pairing speed limits with physical cues to control motion. Cross-shot character consistency requires reusing a fixed character block verbatim; text-to-video and image-to-video prompts are written differently, with the latter focusing on directing motion.

  16. PixVerse 中文官方博客30

    Kling 4.0 vs Seedance 2.5: AI video model comparison

    Kling 4.0 and Seedance 2.5 target demanding AI video workflows but have different emphases: Kling 4.0 focuses on cinematic control, longer shots, and high-resolution delivery, suitable for single-subject cinematic shots and brand key visuals; Seedance 2.5 is designed for multimodal reference packs, suitable for reference-heavy material, multi-subject consistency, and iterative work driven by source material.

  17. PixVerse 中文官方博客38

    Kling 4.0 vs Kling 3.0: detailed breakdown of core differences

    Kling 4.0 extends single-generation duration from Kling 3.0's 15 seconds to 30 seconds, expands keyframes from start/end frames to up to 10, and adds multi-reference Omni Reference control plus stereo audio and improved lip sync. Kling 3.0 remains suitable for cost-sensitive short social media videos under 15 seconds and dialogue-driven clips, while Kling 4.0 targets commercial and narrative content requiring longer shots, multiple beats, and high brand consistency.

  18. PixVerse 中文官方博客22

    Is Kling 4.0 free? Credits, limits, and alternatives

    Kling 4.0 is not free for everyone; it is being rolled out in phases, first opening Kling 4.0 Flash to eligible higher-tier members, while free users still mainly use earlier Kling models. Free accounts typically offer one-time sign-up credits and daily refreshed credits, but differ from paid plans in model access, resolution, duration, audio, watermarks, queues, and commercial usage rights. Free access is better suited for testing and learning; before using it for client delivery or paid marketing, check the commercial use terms of the current plan.

  19. PixVerse 中文官方博客22

    How to use Kling 4.0: a step-by-step guide to AI video generation

    Kling 4.0 can generate short AI videos from text ideas or still images, and the guide provides step-by-step workflows for text-to-video and image-to-video. Prompts are organized in four layers: subject, action, scene, camera and style; first validate the idea with the shortest duration, then generate the final long shot. Image-to-video should focus on describing motion rather than restating the image, and the first version should avoid stacking advanced controls such as lip sync and motion brush.

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

    Why are images blurry when training a realistic character LoRA with Krea 2 on fal.ai?

    A user trained a realistic character LoRA with Krea 2 on fal.ai, using a dataset of 26 images at 1024 resolution generated by Nano Banana and training for 1000 steps. The resulting generations had blurry details such as beards and body hair, while image quality was normal without the LoRA. The user said they had tried various methods without resolving the issue and posted asking for the cause.

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

    Experiment: generating a 720p narrative video with Wan2GP on an RTX 3090

    A creator used Wan2GP with MiniMax H3 Ref2VA Pruned 20B to generate the first episode of the black-and-white sci-fi short drama "The Prompting Zone" on a single RTX 3090 24GB at 720p and 20 steps, with each roughly 5–10 second clip taking about 30 minutes. He first wrote the full script, then generated a 2×2 four-panel composition reference image for each scene as a storyboard, but the model sometimes misread the reference image as a four-panel frame; character consistency and on-screen readable text remain the biggest challenges.