Skip to content
Date and source

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

Oct 17Sat
  1. note · kasmintea · 图像视频创作分享12

    AI illustration 3 specialized in atmosphere: using particles to create a sense of air

    To create a sense of air in AI illustrations, particles can be added after atmosphere and shadows. Placing particles of different sizes in front and behind creates depth, and the particles can be replaced with elements such as light, dust, leaves, water droplets, playing cards, or casino chips to strengthen presence and complete the relationships in the image. As small, numerous elements, particles are more easily perceived as a sense of air rather than noise, but how to eliminate noise remains a challenge.

  2. Frank Houbre · AI电影制作实战50

    Higgsfield 2026 hands-on: what one director kept and what he threw away

    Director Frank Houbre tested Higgsfield 2026: one account can call more than twenty models including Seedance 2.5, Kling 3.0, Veo 3.1, and Sora 2; Cinema Studio 4.0 is based on Seedance 2.5 and provides Director's Panel parameters such as camera, lens, and aperture, with up to 50 references and 30-second clips per generation.

  3. Reddit comfyui · 创作与工作流分享15

    Request for help with MiniMax H3 REF2VA workflow for generating long videos on an RTX3060

    A user is running MiniMax H3 on an RTX3060 (12GB VRAM, 64GB RAM) to generate long videos, using the int8 convrot model and int4 encoder, with memory usage around 50GB and peaks exceeding 60GB. They use the standard ComfyUI template plus VRAM/RAM cleanup nodes, but most REF2VA workflows on Civitai rely on old custom nodes or throw errors, so they are seeking a working workflow that can support long videos with 12GB of VRAM.

  4. Reddit comfyui · 创作与工作流分享12

    Using Flux + Qwen + Architecture LoRA for architecture image-to-image in ComfyUI, detail and image cleanliness insufficient, seeking tuning advice

    A user has built an image-to-image workflow for architectural rendering in ComfyUI using FLUX, Qwen, and Architecture LoRA. The overall direction is right, but the level of detail and image cleanliness are still not ideal. The user is asking the community for advice on adjustments to denoising settings, prompt/conditioning handling, sampling, and related areas.

  5. Monster社区 · 中文ComfyUI实测与工作流22

    Kroma V0.3.1 Turbo: hands-on test of text-to-image startup and ComfyUI entry

    Kroma V0.3.1 Turbo demonstrates the process of launching and entering the default text-to-image interface, which includes a prompt input area, size settings, iteration steps, and a generate button. When using ComfyUI, a custom port must be accessed; the demo mentions 8187, but the actual value should depend on one's own environment configuration. The original video states that this version supports 4-step iteration and is usable with 8G of VRAM; this article has not independently verified VRAM usage or speed.

  6. Reddit comfyui · 创作与工作流分享10

    ComfyUI 0.40.0 workflow crashes after dragging in output with metadata?

    A ComfyUI 0.40.0 user reports that after dragging in an output with metadata, the browser briefly freezes, the entire workflow is blown apart, all subgraphs disappear, and all nodes are flattened onto the main graph; it also swallows the current workflow name, making it easy to overwrite. Rolling back comfyui-frontend-package to 1.52.7 restores stability, and the user suspects rgthree is the main culprit.

  7. Reddit comfyui · 创作与工作流分享10

    How to undervolt a 4080 Super GPU for ComfyUI use

    A user on the Reddit ComfyUI subreddit is asking for help undervolting their 4080 Super with MSI Afterburner, mainly for ComfyUI image generation. After following YouTube tutorials, ComfyUI either crashes outright or outputs all-black images, and nearly all undervolting tutorials they can find online are aimed at gaming, with none related to ComfyUI, so they want to ask whether there are more suitable steps.

  8. 我爱计算机视觉 · 公开公众号 RSS38

    Tencent Photon x NUS Release Orbia Video World Model Evaluation Benchmark

    Tencent Photon and NUS propose the Orbia video world model evaluation benchmark, using 720 exploration and revisit cases to test 27 models, checking retention of known scenes, persistence of newly generated content, and cross-view 3D consistency. SolarWM-H3 ranks first overall with 76.32, followed by Lyra2 and EVOKE in second and third place; content persistence in long video generation is generally lower than in short videos, and the average score for forward-backward trajectories in the world model group drops from 66.21 to 40.96.

  9. Reddit comfyui · 创作与工作流分享12

    minimax H3 Voice Prompts with Audio Reference Fail, Cannot Change Volume or Timbre

    A user reported that in the minimax H3 reference-image-to-video workflow, after inputting about 20 seconds of normal speech as an audio reference, voice prompts produced no change at all, and all outputs had the same volume and timbre. Attempts to use annotations such as says and whispers quietly were ineffective, and changes such as shouting, whispering, or an excited tone could not be achieved.

  10. TelosCompass · note AI视频创作实践22

    Corrected AI Video Should Be Compared with the Pre-Correction Version at the Same Time Point Before Use

    When correcting AI video, keep the pre-correction video and generation conditions, change only one prompt and generate only once, then pause both the pre- and post-correction versions at the same playback moment and compare the same position. The basis for judgment is whether the target problem is fixed, whether new problems appear, and whether the required duration and expression are preserved; then decide whether to use it directly, use only the usable segment, change the approach, or hold off. The article uses a 5-second bike shop video generated by DomoAI as an example, comparing the front wheel shape at 0:03; because the saddle color changes from 0:04 onward, only the 0:00 to 0:03 segment is ultimately used.

  11. Reddit StableDiffusion · 创作者与工作流分享66

    Hands-On with MOSS-Image 5B INT8 Quantization and 8-Step DMD on a 12 GB RTX 3060

    The author tested MOSS-Image (5B single-stream DiT, Qwen3.5-9B text encoder, FLUX.2 VAE, Apache 2.0) on a 12 GB RTX 3060 using 35 difficult prompts and a single seed (42), without cherry-picking images and including failure cases.

    Why it matters: A hands-on test of MOSS-Image's INT8 quantization and 8-step DMD acceleration on 12 GB of VRAM, with a full comparison of text rendering, layout and failure cases.

Oct 16Fri
  1. note · aicreatorpathnet · 图像视频创作分享62

    Using Midjourney 8.2's Edit Model to unify 10 AI character designs

    The author created 10 characters (9 animals plus 1 human) with Midjourney 8.2 for a YouTube AI comedy show, and summarized the core approach to unifying their appearance: use the Edit Model with the exact same two reference images attached each time (a full-body three-view and a face close-up), and state in the prompt what role each reference image plays.

  2. note · promptdesign · 图像视频创作分享18

    A collection of image prompts suited to Instagram ㉘: 100 JSON prompts you can copy directly into ChatGPT, Gemini and more

    This collection compiles 100 JSON-format image generation prompts that can be copied and used directly, aimed at tools such as ChatGPT and Gemini, covering everyday scenes like a gas station at night, a taxi after rain, a station clock, a fishing harbor and a hotel corridor, using cinematic composition and light-and-reflection to create a sense of story. For each subject, besides the base version, it also provides a 9:16 vertical phone version, a 1280×670 note header version, and a variation that changes the background composition and lighting atmosphere.

  3. note · singtothesky79 · 图像视频创作分享15

    Self-taught AI video making ④: the unexpected unfolding from writing a novel with Gemini to manga, animation and a theme song

    A self-taught AI video creator used Gemini to generate a Japanese sword-themed novel in 6 parts, then tried to adapt it into manga and animation. The manga stage yielded only storyboards and images that couldn't form a complete work, but the image quality was very high; after switching to animation, they succeeded in making a "moving manga," and after adding color and voiceover, a theme song was unexpectedly generated automatically.

  4. Reddit StableDiffusion · 创作者与工作流分享45

    FLUX.3 Klein is mediocre at image generation but nearly perfect at consistent photo editing

    FLUX.3 Klein performs mediocrely at image generation but is nearly perfect at high-resolution photo editing that preserves consistency, pose and camera angle. A photographer's hands-on test on an RTX 5090 with 64GB of memory, using flux3_image_klein_sd_bf16 with the qwen3vl_4 text encoder at 3000×4000 resolution and 16 steps, took about 230 seconds to generate a single 12MP image, with no LoRA used throughout.

  5. note · kawakijourney_ai · 图像视频创作分享22

    The complete method for earning 1.2 million yen with YouTube "GROOVE-style BGM" made using AI tools such as Midjourney, ChatGPT and Suno

    Vintage Park, a US YouTube channel registered in December 2025, has accumulated 2,517,047 views, 88 videos and about 14,900 subscribers in roughly 10 months with AI-made GROOVE-style BGM in soul, funk, jazz and hip hop styles, with estimated cumulative ad revenue of about 760,000 to 1.26 million yen at 0.3 to 0.5 yen per view.

  6. note · fordtaurus116 · 图像视频创作分享17

    Setting characters with ChatGPT and drawing illustrations with GPT image2.5: an AI character creation workflow shared

    Using ChatGPT to progressively settle on appearance, personality and manner of speech, then using GPT image2.5 to draw the settings as illustrations, is the core of this character creation workflow. Using the original character Apollo as an example, the author first settles on blond hair, emerald eyes, tan skin and yaeba teeth, then derives expressions and poses from his harmless yet slightly devilish personality. Finally, the full body, expressions and front/side/back three-view are combined into a single character sheet, and the outfit is changed from casual clothes to a gorgeous kimono to verify how the same character performs in different outfits and scenes.

  7. note · promptdesign · 图像视频创作分享22

    Hand-drawn daily illustration image prompt collection ⑱: 100 prompts ready to copy into ChatGPT and Gemini

    This prompt collection contains about 100 "gentle daily life" hand-drawn illustration prompts in JSON format that can be copied directly into ChatGPT and Gemini. In addition to the basic prompt, each subject also provides a 9:16 vertical version, a 1280×670 note header version, and two variants that change the background, lighting, clothing and time of day. Subjects cover family, seasons, street scenes and everyday life, suitable for note headers, SNS posts, blog title images and advertising materials.

  8. note · myloose · 图像视频创作分享8

    [AI portrait generation research] Generating a monolid man with the top-tier model Astra ⑧ "Superstar, Fallen"

    Using the top-tier model 6astra to generate monolid male illustrations with the same prompt, the first image succeeded on the first try, but after adding a shiny art style the double-eyelid lines were reinforced and the profile turned completely into double eyelids. When attempting local specification, the eyebrows were erased, and even when specifying monolids per the character sheet and generating two people in daily life, although the atmosphere and finish were good, a monolid could never be generated.

  9. Promptsref.com · 创作分享15

    Midjourney Top 3 SREF of the Week: Healing-Style 3D, Ink-Wash Grotesque, and Cool-Tone Fashion

    Promptsref announced this week's top three Midjourney SREF styles: first place is a dual-code hybrid healing-style 3D style that combines Pixar-like character charm with children's picture-book colors and toy-grade 3D texture, suited to mascots and parent-child visuals; second place is a single-code ink-wash grotesque style that blends Chinese brushstrokes, rice-paper texture, and exaggerated contemporary character design; third place is a single-code cool-tone fashion style that uses hard flash and high-fashion styling to create a magazine-editorial feel.

  10. note · kt_kt_89 · 图像视频创作分享22

    After Building My Own Video Generation Pipeline, I Realized Runway Workflow Already Existed

    A developer spent a long vacation building his own video generation pipeline, only to discover afterward that Runway's Workflow feature could already chain inputs and multiple generation steps into a reusable process, quite close to his own approach. He tested it with a 3D short animation featuring a cat protagonist; the process ran from theme, to three-card storyboard, to each card's starting image, to video generation, to merging, but adjustments still produced deviations such as leftover inputs, aspect-ratio reverting, changed character descriptions, and missing model settings, requiring values to be checked directly in the editor.

  11. Hugging Face Daily Papers17

    Gen4P: The Video Generator Is a 4D Perceiver

    Gen4P proposes a unified framework that can turn a pretrained video generator into a feed-forward 4D perceiver; without iterative denoising, a single encoder forward pass can process clean video latents and predict the required 4D attributes. Its on-demand querying supports flexible spatial sampling, continuous-time queries provide temporal flexibility, and a pretrained video-depth LoRA can be plugged in as a frozen geometric prior. Gen4P achieves SOTA or competitive performance on 3D point tracking, video depth, and camera pose estimation tasks.

  12. note · hagiri1312 · 图像视频创作分享22

    Using Grok to build an image prompt selector: adding "lock" and "reroll only this item"

    The author had Grok build a personal illustration-generation prompt selector that assembles instructions to pass to image generation using Japanese options, and does not generate images itself. The initial version on October 15 randomly rerolled all 5 items—character, clothing, pose, background, and scene—together. After reading the HTML, Grok proposed an improvement: "lock the items already decided and reroll only the rest." The revised HTML has added per-item "lock" and "reroll only this item," and hairstyle and hair color have also been split into independent selections, but the author only confirmed the code level and did not test all buttons.

  13. note · kiryu_mizuki · 图像视频创作分享22

    Running AI image generation on a 6GB VRAM laptop: INT8 practice with Krea2 Turbo and Qwen-image2.1

    Creator Shuihui runs AI image generation locally on an RTX 3060 Laptop (6GB VRAM), generally choosing INT8 versions of models rather than GGUF versions, considering the latter's image quality not to their taste. Currently, generation with Krea2 Turbo plus 4STEPLORA takes about 70 seconds, Qwen-image2.1 with the official int8 model and 8STEPLORA takes about 65 seconds, and two reference images take about 70–80 seconds.