Skip to content
Date and source

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

Oct 20Tue
  1. Reddit StableDiffusion · 创作者与工作流分享22

    Why does training complex styles with MOSS always fall short? User asks for help after nine failed attempts

    A user trained a complex style with MOSS, but the result was only superficially similar and clearly different from the original style; after nine attempts with different caption styles and training parameters, it still did not meet the target. After raising the learning rate to 5 times the commonly recommended value, the generations showed broken human anatomy, while the style was no different from that at the low learning rate. The user is publicly asking for parameter-tuning experience.

  2. TelosCompass · note AI视频创作实践38

    Generating PR video storyboards with ChatGPT: revising the AI plan scene by scene

    When producing a 15–30 second PR video, first submit the subject, audience, core value, and format conditions to ChatGPT all at once to generate a storyboard plan, then revise each scene by rewriting atmospheric words like "soft rainy scene" into "what is seen, what moves." In the example, a 20-second video is split into four 5-second scenes, with the protagonist fixed as a dark blue long umbrella, and people, dialogue, and hand movements all removed; the final scene reuses the previous scene's footage with text added. Before use, check the four items of purpose, facts, production, and duration, and keep only one version of the planning memo.

  3. note · kongo_jun · 图像视频创作分享55

    Testing Qwen-Image-2.1's Multiple-Angles LoRA in ComfyUI

    The author set up and tested Qwen-Image-2.1's Multiple-Angles LoRA in ComfyUI, which can change the angle of the photographed subject via prompts. In testing, a single image took 23 seconds to generate and used 21GB of VRAM, but this LoRA currently focuses only on subject angle changes, the background disappears, and the output is fairly random. The article provides the custom node installation command, the placement paths for the model and LoRA files, and the workflow file.

  4. Reddit comfyui · 创作与工作流分享72

    Flux 3 Klein tested in ComfyUI: 19 free workflows covering generation, editing, and compositing

    The author tested Black Forest Labs' new image model FLUX 3 Klein in ComfyUI, with 19 accompanying free workflows covering text-to-image, single-image editing, and multi-image compositing.

    Why it matters: The author tested Flux 3 Klein's 19 free workflows in ComfyUI and listed the model's shortcomings in hands, faces, and text.

  5. note · synth_brain · 图像视频创作分享75

    Making a MiniMax H3 continuous short drama in ComfyUI with obvpm-timeline and RefMod

    The author introduces a new combined workflow for MiniMax H3 in ComfyUI, using the lightweight timeline node obvpm-timeline together with RefMod for fixed characters to complete generation, backward extension, forward insertion, and split insertion within the canvas. The article says the timeline connects directly to the latent data of each shot, avoiding the quality loss and audio-video misalignment caused by MP4 concatenation, and includes links to the workflow, Colab launch code, and node repository.

    Why it matters: The author tested combining obvpm-timeline with RefMod, demonstrating continuous short drama production with forward/backward extension and split insertion within ComfyUI.

  6. note · singtothesky79 · 图像视频创作分享20

    Making and releasing an animated video with AI only, part 2: from story to characters to the finished film

    The author completed the opening portion of episode 1 of an animation using AI only: first writing the story, then creating the characters, and finally generating the video, with the finished product showing a clear animation feel. Previously, a video of their AI character "シーサーわん" had gained some views, and this time they expanded it into a finished film based on the introductory segment of animation episode 1. The author plans to continue releasing sequels and gradually complete the entire storyline.

  7. note · kamiryu_ai · 图像视频创作分享48

    Before making a character ad with AI video, I first made a character reference sheet, but a person still appeared in the mirror

    To make a roughly 14-second vertical beauty salon ad, the author first used Codex to generate a character reference sheet (left side: front/side/three-quarter face; right side: full-body front/side/back), then handed it to Higgsfield to generate the video in one pass, taking about 5 hours in total. The faces in the finished film remained largely consistent, but a figure in a white shirt still appeared in the mirror from 1.9 to 2.6 seconds, even though the prompt explicitly said it should not appear; in the end, a 0.667-second segment was simply cut, leaving a finished film of about 14.4 seconds, with no additional credits consumed.

  8. Reddit comfyui · 创作与工作流分享60

    The author shares a ComfyUI guide and workflow for H3 Minimax body and background replacement without LoRA

    Reddit user Dudelydad78 shared a guide and runnable example for replacing the body and background with H3 Minimax in ComfyUI, saying no LoRA is needed. The materials also include the full guide PDF, a comparison of the original and generated videos, and their personal workflow Ironmax.json containing their custom nodes. The related files are all hosted in the BaloneyKit for Minimax H3 directory on Hugging Face.

  9. note · promptdesign · 图像视频创作分享22

    New AI beauty image prompt collection ⑰: 100 JSON templates that can be copied directly into ChatGPT and Gemini

    A prompt collection for AI beauty image generation, containing about 100 JSON-format prompts themed around travel-photography scenes from around the world and Japan, which can be copied directly into tools such as ChatGPT and Gemini. Each theme provides a basic version, a 9:16 vertical version, a 1280×670 horizontal note-cover version, and background and atmosphere adaptation versions, along with negative_prompt and aspect_ratio fields.

  10. note · ponko_ai2026 · 图像视频创作分享22

    Is this composition okay? Learning camera movement and perspective with ChatGPT | AI Image Creation Practice ②

    In the second installment of the AI image creation series, the author uses ChatGPT to learn camera movement and perspective, sorts out the three elements of camera height, distance, and orientation, and distinguishes the expressive effects of low angle, close distance, and Dutch angle (tilted composition). The composition is then designed as "a low camera position close to the ground, slightly looking up, frontal shooting, highlighting the right hand holding magic," with ChatGPT completing the body twist and the division of roles among the limbs and generating the prompt. The first generated image has a strong perspective impact but the hands are somewhat thick; after regenerating with the same prompt, the hand proportions return to normal.

  11. Reddit StableDiffusion · 创作者与工作流分享70

    How to train a MOSS-Image LoRA locally on 8 GB of VRAM: parameters, tips, and workflow

    The author disclosed a set of parameters for training a MOSS-Image character LoRA locally on 8 GB of VRAM: the default setting trains for 800 steps, extended to 1200–1500 steps for large datasets or difficult concepts, with a learning rate of 0.0004, rank/alpha 8/8, and 512² resolution.

    Why it matters: The author provides the complete training parameters and step range for a MOSS-Image character LoRA, and explains why 8000 steps would overfit, serving as a starting reference for local training.

  12. Higgsfield 官方博客22

    How to make AI real estate videos from photos: a hands-on guide to Higgsfield Cinema Studio 4.0

    Higgsfield released a tutorial on how to use Cinema Studio 4.0 to turn real estate photos into video clips of each space one by one, then stitch them into a property tour with listing information and music. The workflow includes uploading clear photos as references, setting camera/lens/aperture/lighting/pacing in the Director's Panel, choosing one gentle movement per clip from 30+ camera movement presets, and using Extend and Edit Video to fix the results.

  13. Behind the Verticals · 竖屏剧制作简报70

    VX Studio founder's retrospective: using AI video to make a 90-episode vertical drama, which parts break down

    VX Studio founder Shaurya Gupta uses a 90-episode vertical drama with about 1600 generations as an example to explain that AI video engines have no memory, so the same character across shots shows drift in face shape, age, and height, and by the fourth consecutive extension it already looks like a different actor.

    Why it matters: The founder of an AI short-drama production platform reviews hands-on experience from a 90-episode vertical drama, listing point by point the failure points of character consistency, shot extension, and wide-shot distortion, along with countermeasures.

  14. note · nobinlog · 图像视频创作分享58

    Using the lightweight Qwen-image2.1 in ComfyUI for three black-and-white line art workflows

    In ComfyUI, the author uses the lightweight Qwen-image2.1 together with Flux2's flux2-klein-4b-fp8 high-speed version to produce three black-and-white line art effects: fine-line-preserving, solid black, and comic style, runnable on 8GB of VRAM. The author says that for simple conversion to black-and-white line art, Flux2 is faster, but the difference between the two results is not obvious; the workflow automatically generates eyes for faceless characters, which needs to be adjusted with prompts.

  15. Oimi · 中文图像视频教程76

    Getting started with MOSS-Image in ComfyUI: installation, sampling parameters, and JSON layout control

    The OpenMOSS team open-sourced the 5B text-to-image model MOSS-Image on October 16, 2026, with weights under Apache 2.0, and it was supported in ComfyUI on the day of release.

    Why it matters: A review of MOSS-Image's installation, sampling parameters, and VRAM trade-offs in ComfyUI, explaining the coordinate conventions and current boundaries of JSON layout control.

  16. 可灵官方博客38

    20 video ideas for Kling 4.0: lifestyle, tutorial, and story topics for content creators

    Kling 4.0 supports text-to-video and image-to-video, with a single clip up to 30 seconds long, maintaining more stable imagery during large movements and complex camera work. Its Omni Reference can combine up to 15 image, video, and sound references plus Elements, and can additionally use up to 10 uploaded images as keyframes to control shot nodes. Officially, 20 video idea examples are provided around directions such as daily life, tutorials, personal stories, hobbies, and festivals.

  17. note · kiryu_mizuki · 图像视频创作分享30

    Mizuki shares three ways to write AI image generation prompts: single words, fixed templates, and natural prose

    Mizuki shares three ways to write AI image generation prompts on an RTX 3060 Laptop (VRAM 6GB): listing single words has high randomness; fixed templates with itemized instructions by style/composition/pose, etc., are easy to correct; natural prose is the most commonly used. All three methods also work on generative AI such as GPT and Gemini. The author says recent models sometimes do not write negative prompts, and writing them may instead cause them to be generated.

  18. note · aun_585 · 图像视频创作分享58

    Full storyboard prompts for the Seedance 2.5 first-person animated short "Dinner After Overtime" released

    The author has released all the video prompts for the first-person animated short "Dinner After Overtime" made with Seedance 2.5. The process is to first use Midjourney to generate the first and last images of each shot, then use Seedance 2.5's keyframe feature to animate between the two images.

  19. Ciaro · 中文AI电影制作教程22

    Multi-format video editing: how to make approvals for multi-version edits traceable throughout

    The key to traceable approvals for multi-format video editing is to first formulate and confirm a delivery matrix before editing, recording item by item the distribution channels, formats, end cards and subtitles, audio files, and the review owner and final approver. Each review version corresponds to only one clear pending decision and is labeled with version, date, and format, for example COM-16x9-v07 · October 10, 2026 · pacing and shot order. Client approval of creative content does not equal confirmation of the final delivery files; before delivery, the source version, cropping, subtitles, mix status, and export specifications must also be verified item by item.

  20. Frank Houbre · AI电影制作实战67

    How to keep the same character consistent across shots in AI video

    Frank Houbre gives a production method for keeping the same character across shots: first make a front-facing main image plus two or three angle references, then use the model's built-in character reference entry point rather than pure text prompts, and finally break down the storyboard by shot risk.

    Why it matters: The author horizontally compares the character reference entry points of Kling, Seedance, Flow, and Runway according to official documentation, and provides a reusable storyboard and reference image preparation process.

  21. note · aiholic0831 · 图像视频创作分享40

    How big is the difference between 1K and 2K resolution in image generation AI? Upscaling comparison with Grok using the same prompt

    Using the Grok API's grok-imagine-image-quality model at a 3:4 ratio with the same prompt to generate 1K (864×1152) and 2K (1776×2368) images for comparison, it was found that 1K is already sufficient for close-up faces, while the difference is obvious in compositions where the person is smaller, such as full-body shots: at 1K eyelashes blur into line blocks and pearl earrings become blurry dots, while 2K can draw eyelashes individually and retain highlights.

  22. note · quirky_hippo6791 · 图像视频创作分享22

    Trying AI comics with ChatGPT: an experiment with four ways of placing short text and comics

    The author conducted four sets of AI comic generation experiments with ChatGPT, with the core variable being the placement of comics relative to short text: turning an article into a one-page comic, pairing one page of comic before and after an article, inserting comics in the middle of the main text, and using a "novel—four-page comic—novel" structure to let the comic carry mid-length narrative. The experiments also specified art materials such as transparent watercolor, Copic, and liquid ink, and emphasized that the intent of the surrounding context must be explained to ChatGPT, otherwise the differences between placement methods will be flattened out.

  23. Reddit comfyui · 创作与工作流分享20

    Minimax H3 ref2va cannot replace video characters, user seeks help with prompts and workflow

    When using Minimax H3 ref2va (DaSiWa) for video character replacement, the model over-retains the original video content and only replaces parts that do not appear in the frame, failing to complete identity replacement. The user asks whether ControlNet is needed and seeks usable prompts or workflows.

  24. note · freelife_creator · 图像视频创作分享38

    How to solve "all AI portraits look the same": 5 ways to establish a unique visual style

    In response to the common problem that images generated by tools such as Midjourney and DALL-E 3 have a strong "AI flavor," a personal visual style can be established through 5 prompting methods: mixing two or more styles (such as ukiyo-e × steampunk), using feature descriptions such as "swirling brushstrokes" instead of artist names to avoid copyright risks, locking in a 3-color brand palette, specifying textures such as fabric/paper, and fixing 3-4 composition patterns.

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

    Generation stuck at 33% after ComfyUI update: user seeks help with foxydits workflow slowdown

    After a user updated ComfyUI on October 12, a 12-step 1.0mp generation task that originally completed in 20-26 minutes slowed to 2 hours, with the sampler stuck at 33%. The configuration is a Ryzen 7 5700x, 5070Ti 16GB VRAM, 48GB RAM, using the foxydits seedhunter v2.5.1 workflow with the singularity model plus turbo lora.

  26. Reddit comfyui · 创作与工作流分享12

    User's first attempt at running the Wan motion transfer template with ComfyUI 0.42.1 portable on a Legion Pro 7 fails

    A user tried the ComfyUI 0.42.1 portable version for the first time, chose the Wan motion transfer template, and after installing all models, running it failed immediately: either it disconnected directly with no error, or it failed at the sampler stage, and the same happened with the regular version. The device is a Legion Pro 7 with 32GB RAM and a 5090 24GB VRAM, the NVIDIA driver has been updated, and the user is unsure whether they should run the 3 update bat files.

  27. Hugging Face Daily Papers45

    Manga2Animate: An agent framework for faithful manga animation using script and storyboard planning

    Manga2Animate is a multi-agent framework that breaks manga animation down into two intermediate representations: an animation script and a visual storyboard. Each representation is generated by a creation agent and reviewed by a paired review agent against the original manga storyboard. The framework uses representation-specific selection and a two-layer feedback mechanism to progressively refine both, then synthesizes the final animation. Experiments and user studies show that its output is more faithful to the original manga and has better shot-to-shot consistency than existing state-of-the-art methods.

  28. Hugging Face Daily Papers45

    Stateful Harness: A world model framework that separates rule execution from video rendering

    Researchers propose the Stateful Harness framework, which separates world state evolution from video rendering: rules maintain the world state, a frozen video model renders each event, and the generated clips are then used to update the state. The team built WorldState-Bench, containing 50 pairs of counterfactual twin samples; with rule prompts alone, the best of 11 video backends was only 32% correct; MiniMax-H3 rose to 64% when the ending was explicitly specified, and reached 92% with the full framework connected.