Skip to content
Date and source

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

Oct 23Fri
  1. 可灵 Kling58

    Kling uses Kling 4.0 to make the KING CARD short film, breaking down dynamic motion and shot scheduling

    Kling releases behind-the-scenes breakdown of KING CARD, explaining how this short film, which has accumulated 5 million views on social platforms, used Kling 4.0 to complete multiple transitions from playing cards to a city and a starry sky. The article mentions that Dynamic Motion handles complex movements such as the card spiral, multi-keyframe control handles key nodes such as cards turning into paper birds, Omni Reference maintains visual continuity, and high-angle wide shots, lateral tracking shots and FPV perspectives are used to convey the height and speed of the newspaper train. Kling 4.0 is now fully open.

  2. Monster社区 · 中文ComfyUI实测与工作流52

    Singularity-V2 Bundle Demo: 4-Step Video Generation, First/Last Frames, and ComfyUI Entry Point

    This tutorial compiles a short demo of the Singularity-V2 bundle from Liu Yue's tech blog, explaining that the bundle defaults to 4-step iteration. The short video actually demonstrates three input methods: text-to-video, first/last frames, and all-purpose reference. The author states the demo device is an RTX 4060 laptop, with a single generation taking approximately 100 seconds; this duration corresponds only to the conditions of this demo. To view node connections, you can access the ComfyUI workflow via custom port 8187; the port should be confirmed from the local startup log.

  3. Reddit comfyui · 创作与工作流分享22

    Krea2 edit node loses detail: user seeks a way to preserve source image details 1:1

    A user reports that although Krea2 is not an editing model, operating through the edit node causes the output to lose a large amount of detail and exhibit pixel drift. They have tried various sampler/scheduler combinations, 8/10/12 steps, secondary upscaling, Krea2 enhance LoRA, raw/turbo models and the Wan2.1 VAE, all without noticeable improvement, and ask whether they can only wait for the local version of Krea3.

  4. Reddit StableDiffusion · 创作者与工作流分享15

    How to add multiple LoRAs to the Minimax H3 default t2v workflow

    A user on Reddit asked how to add additional LoRA nodes to the default ComfyUI text-to-video workflow of Minimax H3. The default workflow comes with a turbo LoRA and multiple switch nodes; the user said they can already configure multi-GPU operation but cannot understand the switch logic, and LoRA is simply ignored during actual runs.

  5. Krea 官方博客62

    How Krea uses Krea: from a layered tech stack to a self-trained agent team.

    The Krea team used its own tools to produce the Krea 3 launch video and benchmark video, cutting costs from over $200,000 to about $300 and time from 2 months to within 24 hours. The workflow began with Figma storyboards, went through Krea MCP, Claude, and After Effects plugins, and was ultimately completed by the self-trained agent Wobble, which produced 116 versions of the benchmark video.

    Why it matters: The Krea team has published a layered workflow from Figma storyboards to self-trained agents, along with prompts and feedback mechanisms, so creators can adapt this approach.

  6. Reddit comfyui · 创作与工作流分享67

    Author tests AI character replacement with MiniMax H3 in ComfyUI and releases a workflow tutorial

    The author tried MiniMax H3's Reference-to-Video workflow in ComfyUI, using AI-generated characters to replace people in the original video while preserving the original footage's action and performance, and also attempted to replace the background environment.

    Why it matters: The author tested MiniMax H3's reference-to-video process locally, organizing character replacement and motion transfer into downloadable workflows and assets.

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

    Seeking help with MOSS-style LoRA training: 150 images, 30 epochs, and loss stuck at 0.44

    A beginner trained a MOSS-style LoRA with FizGig's default settings, carefully selecting and strictly captioning 150 images and setting 30 epochs, but by epoch 18 the loss could not go below 0.44 and kept rising back between 0.48 and 0.56, so they stopped training and asked the community whether to continue or adjust parameters.

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

    How do you avoid flicker or colored dots when extending video with Minimax H3 in ComfyUI?

    When extending video with Minimax H3 in ComfyUI, the user saw flicker at the start of the second segment and colored dots throughout the whole segment; they tried adjusting context frames, window 39/90/141, snap cuts, quality, and aspect ratio, and switched between the ref2va and fl2va models, all to no effect.

  9. Reddit comfyui · 创作与工作流分享22

    ComfyUI generation speed improves 20%-100% after swapping the RTX Pro 6000 power cable: user's measured Wan 2.1 and Flux 1d time comparison

    After a user replaced the CableMod 12V-2x6 cable (4 PCIe connectors) on an RTX Pro 6000 with a Corsair thermal-protection cable, Flux 1d generation time in ComfyUI increased by 20%, and Wan 2.1 720p went from 78.42 seconds to 180 seconds. After the cable swap, TDP dropped from a peak of 105% to about 90% and board power from 650W to 610W; the user is considering switching back to the old cable.

  10. Reddit StableDiffusion · 创作者与工作流分享58

    Creator shares Flux3 style transfer workflow with adjustable composition and style strength

    Reddit user Winter_unmuted shared a one-shot, adjustable image-to-image style transfer workflow based on Flux3, centered on a reference latent plus node. The author says composition strength (the later the start, the weaker and more creative) and the strength of the image style relative to the prompt (the earlier, the more image style; the later, the more prompt-dependent) can be adjusted separately, and includes gradient examples.

  11. Kapwing 创作指南58

    Kapwing tutorial: how to translate subtitles and dub videos with Claude

    Kapwing released a tutorial explaining that Claude, connected to 48 tools via Kapwing MCP, can complete three workflows: subtitle generation, subtitle translation and AI dubbing. The author tested with a 45-second two-person podcast and a 21-second talking-head video; for the latter, five languages were requested in one go, and five dubbed versions were generated in 76 seconds.

  12. Reddit comfyui · 创作与工作流分享20

    How to identify duplicate models and LoRAs on a hard drive?

    While cleaning up their hard drive, a user found multiple models and LoRAs with similar names and exactly the same file size, such as fp8mixed.safetensor and fp8.safetensor, and suspected duplicate files. They asked whether there is a way to directly determine whether these models or LoRAs are truly duplicates, so one can be deleted to free up space, or whether the only option is to run them separately in ComfyUI workflows and compare the outputs before deciding what to delete.

  13. note · aoao_ai · 图像视频创作分享42

    Drawing the same character continuously with AI: my homemade reference sheet and Supabase management method

    The author uses a reference sheet (three-view drawings, an expression list, clothing, and frequently used small props) as reference images to make an image-generation AI keep the same character's appearance across multiple images. They centralize the reference sheet images and information on "which version is the current version" in Supabase, with statuses divided into current, old version, and returned/in production, and use only the current version when drawing. Even so, breakdowns still occur—for example, curly hair lines becoming faint and eye positions shifting; indicating fixes with numeric pixels did not work, and only attaching example images fixed it.

Oct 22Thu
  1. Reddit comfyui · 创作与工作流分享58

    Pure prompt vs. layout JSON comparison: 8 seeds each for posters, menus, and product labels

    In ComfyUI, the author ran MOSS-Image-SFT with 8 seeds each for posters, price signs, and product labels, comparing text placement performance between pure text prompts and layout JSON. On the five rows of price signs, layout JSON placed all 88 text blocks, while pure text prompts placed 67/88; for posters and labels, the gap between the two was within 1 block; spelling was fully correct for both methods.

  2. TelosCompass · note AI视频创作实践58

    Before using AI to create video assets, first prepare front and side reference images of the product

    Using a fictional umbrella shop "Shizukudo"'s 20-second vertical video as an example, the article explains that before generating asset images shot by shot, you should first use an image generation AI to create reference images of the same product from various orientations, such as front and side, and then attach the reference image to generate each asset image one by one. The author recommends comparing each asset image side by side with the reference image and the first shot's asset image, correcting inconsistencies item by item in text, and confirming there are no frames, numbers, or explanatory text before handing it over to the video generation AI.

  3. Reddit StableDiffusion · 创作者与工作流分享48

    Using LatentOperationSharpen to Sharpen Latents and Mitigate Blurry MOSS Image Output

    To address the blurry output issue of MOSS Image, a two-stage denoising process can be used: first denoise normally for 5 steps, then apply LatentOperationSharpen to the latent image, and then continue denoising. The author used the base model plus LoRA and showed comparison results of stacking the SFT model as a LoRA onto the base model during the second-stage denoising, as well as comparison results of only sharpening the latents.

  4. Reddit StableDiffusion · 创作者与工作流分享60

    What Are the 5GB and 9GB Files in the New ComfyUI 5B DiT Repackaging?

    Based on the Hugging Face file list and official documentation, the author outlines the purpose of each file in the ComfyUI repackaging of MOSS-Image: the 9GB version (10.16 GB on HF, 9.47 GiB on disk) is bf16, and the 5GB version (5.41 GB) is int8 ConvRot quantization; Base and SFT each have two versions.

  5. note · ponko_ai2026 · 图像视频创作分享22

    Learning Human Anatomy and Skeleton with ChatGPT: Pose Correction in AI Image Generation

    In the third installment of her AI image creation series, the author uses ChatGPT to help learn the relationships among the ribcage, pelvis, spine, and shoulder-arm joints, then writes prompts accordingly to generate two full-body action images: a roundhouse kick and a stepping-forward pose. She specifies the ribcage and pelvis as separate three-dimensional forms with their own directions and lets the pelvis rotate naturally with the kick direction, making the connection between the hip joints and knees more natural than before.

  6. Recraft 官方博客22

    How to Make Twitch Emotes: Size Specifications and Recraft Studio Workflow

    Twitch emotes should be designed at the smallest chat size of 28×28 and uploaded as square PNGs with transparent backgrounds; static images are limited to 1 MB, and animated images must not exceed 60 frames. Recraft Studio can first generate the initial emote using vector output, then use V4 Styles with up to 10 reference images to lock in a consistent style, and remove the background and export 112/56/28 sizes in one click. Custom emotes are available only to Twitch Partners and Affiliates and must pass platform review.

  7. note · promptdesign · 图像视频创作分享22

    New AI Beauty Image Prompt Collection ⑱: 100 JSON prompts that can be copied directly into ChatGPT and Gemini

    An AI beauty image prompt collection for tools such as ChatGPT and Gemini, containing about 100 JSON-format prompts that can be copied directly, with themes combining spectacular scenery from around the world with adult women. Each theme provides a 1:1 basic version, a 9:16 vertical version, a 1280×670 note image-output version, and background and atmosphere adaptation versions, along with negative_prompt and aspect_ratio fields.

  8. note · ya_ma_da_pon · 图像视频创作分享20

    ChatGPT image generation: Halloween lover AI boyfriend CG prompt

    The author used the ChatGPT image generation model to create a Halloween-themed AI boyfriend CG and provided the complete prompt for "Halloween Lover," requiring a 3:2 horizontal composition and a slightly high-angle overhead view, with the male character retaining the facial features and physique of the reference image but with eye color changed to deep red, and a pumpkin-headed female monster on the right, with beauty and eeriness in a 7:3 ratio. The prompt needs to be used together with the character's "face close-up, standing illustration, three-view" trio to maintain consistency in clothing, hairstyle, and appearance when changing poses and compositions.

  9. note · comfylab · 图像视频创作分享60

    Krea 3's Ultra mode has been merged into the ComfyUI template; the author tested its editing capabilities via API

    Krea 3's Ultra mode was merged into the ComfyUI template on October 22, and the author tested the model's editing capabilities via the ComfyUI API. Krea 3 supports native 2K editing, region editing that redraws only changed areas, compositing of up to 8 reference images, HEX color value specification, and layout text preservation; Ultra has a fixed 2K output and takes several minutes per image.

  10. Reddit comfyui · 创作与工作流分享20

    Eyes are distorted when generating full-body images with Z_image_turbo_bf16; how to improve

    When using Z_image_turbo_bf16 to generate full-body portraits, the user found that facial close-up details were excellent, but the eyes appeared distorted and asymmetrical. After trying the two upscaling models RealESRGAN_x4 and 4x-ultrasharp, the results remained the same, and the positive prompt already included "Detailed eyes, sharp focus, detailed iris." The user asks whether there is a way to improve eye quality in full-body images.

  11. note · kawakijourney_ai · 图像视频创作分享30

    Reproducing the process of an office lady going to a hotel after work to change into Halloween cosplay using Midjourney/ChatGPT/Gemini (with prompts)

    A Japanese creator has released a set of prompts for reproducing the full process of "an office lady walking into a hotel after work and changing into Halloween cosplay" using Midjourney, ChatGPT, Gemini, and other AI generation tools, broken down from STEP1 to STEP5: walking out of the office building wearing an employee badge, walking on a night street, taking off shoes in the hotel entrance, and cosplay poses on the sofa and bed.

  12. note · tsugiito_design · 图像视频创作分享58

    Creating a character with hands pressed against a glass window in ComfyUI: tags and JSON for raindrops, window frame, and reflections

    The author used Animagine XL 4.0 Opt to separately generate the character and raindrop assets, then in ComfyUI used a mask of the transparent area of the window frame to composite the character and raindrops onto the inside of the glass, preserving the front-to-back relationship of the window frame being in front of the character.