Skip to content
Date and source

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

Oct 21Wed
  1. 可灵官方博客33

    What Is Lip Sync? Meaning and an Analysis of Lip-Sync Technology in AI Video

    Kling AI's Lip Sync tool can add speech or singing to an existing character video and align the lip movements, while Kling 4.0 can simultaneously produce dialogue, facial performance, and sound when generating new scenes. Compared with the previous generation, Kling 4.0 improves lip sync, supports more flexible voice references, expands multilingual dialogue, and adds dual-channel stereo, with audio and visuals generated in sync by default.

  2. 可灵官方博客25

    How to Edit Photos with Kling IMAGE 3.0 Omni: A Beginner's Guide

    This guide first covers the three steps of traditional photo editing (cropping and level correction; adjusting exposure/shadows/highlights/white balance/contrast/saturation; exporting JPEG/PNG/WebP), then explains how AI photo editing uses natural-language descriptions to replace objects, change outfits, change backgrounds, and change styles. Kling IMAGE 3.0 Omni supports image-to-image workflows, can combine up to 10 reference images, and natively outputs 2K and 4K.

  3. Hugging Face Daily Papers31

    OverSeer: improving multi-agent world model consistency with a fixed auxiliary camera viewpoint

    OverSeer is a multi-agent world model that, by introducing a spatially fixed auxiliary camera viewpoint beyond the agents' first-person views as a proxy for a world-centric representation, helps the model maintain temporal and cross-agent consistency. The auxiliary view encodes only changes relative to a static reference image via delta-VAE, significantly reducing the tokens and computation required for processing. Experiments on Minecraft, Doom, and real multi-camera captures show improvements in both world and agent consistency.

  4. Hugging Face Daily Papers38

    HRoPE: homography relative position encoding

    HRoPE proposes surface-aware relative position encoding, representing each patch's local geometry as a tilted 3D plane and injecting the plane-induced homography as a relative transformation into attention, to correct perspective distortion and align patch features across views.

  5. note · dreammii · 图像视频创作分享16

    Don't just finish your AI music and leave it there! A 3-step method for a 30-second "teaser short video"

    After generating music with AI, you can turn the work into a 30-second teaser short video in 3 steps: first pick the 30-second segment you most want people to hear, then pair it with 1 image that fits the style and brief text, and finally check volume, text readability, and material copyright before publishing. For submitting to YouTube Shorts, a 9:16 vertical format is recommended, and AI-generated tracks also require confirming the commercial use scope according to the terms of the service used.

  6. Reddit comfyui · 创作与工作流分享54

    Hands-on test and workflow for character replacement using MiniMax H3 reference-to-video plus LoRA

    The author spent about a day testing character replacement using MiniMax H3 reference-to-video with LoRA, and presents unselected early results. The method replaces the person in a video with a character reference image while preserving the original video's motion, scene and audio; the author also lists limitations such as similarity drift, details blending with the original person, and temporal misalignment caused by scene cuts, and says additional prompts help improve consistency.

  7. fal 官方博客28

    EAGLE-4 technical deep dive: Fireworks and fal release speculative decoding benchmark dataset

    Fireworks and fal jointly released a technical analysis of EAGLE-4 and Prompt Enhancement SpecBench, a speculative decoding benchmark dataset for creative tasks. The dataset contains 400 samples, covering 100 each of text-to-video, image-to-video, text-to-image and image-to-image. On Qwen 3.8 27B, EAGLE-4 uses sliding window attention and a low-rank candidate head to reduce draft cost, and the verification tree dynamically adjusts width and depth based on GPU latency measurements.

  8. Loomance · 本地AI动画实测69

    Wan 2.2 comparative evaluation under Loomance and WanGP: Fast Mode and full workflow yield different results.

    The author compared the rendering results of Wan 2.2 under two tools, Loomance and WanGP, and found that the 4-step lightning mode is about 10 times faster but has more errors, while the 30-step full sampling is cleaner. In the six-dog scene, WanGP had 3/5 instances of character melting or merging, while Loomance had 0/5; in high-resolution single-character scenes, the two are close. Choose as needed: use WanGP for rapid changes or low VRAM, use full sampling to maintain character consistency.

    Why it matters: A practical comparison of the same Wan 2.2 model under two local tools offers a reference for creators to choose between speed and fidelity.

  9. Reddit StableDiffusion · 创作者与工作流分享17

    Memory issues when loading LoRA in Krea 2

    A user encountered memory issues when loading LoRA to generate images in Krea 2: with a 5070 Ti GPU and 16GB of RAM, generating one image after loading LoRA took 10 minutes, after which the computer froze. The user is using Forge Neo with a Q8 quantized model, suspects insufficient memory, and is seeking advice on optimizing memory management.

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

    ChatGPT image generation: Aizawa releases free prompts for 5 scenes—morning, day, evening, night and late night

    Aizawa posted 5 illustrations on X from the morning of October 20 to late night on the 21st, and released the corresponding ChatGPT image generation prompts for free, covering five scenes: stretching in the morning, embroidering in profile during the day, waving in the evening, looking back at night, and falling asleep late at night.

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

    Moss-Image SFT with the new DMD2 LoRA: comparison of 8-step generation results

    User AgeNo5351 paired the Moss-Image SFT model with the new DMD2 LoRA to generate images in 8 steps, and compared the results with old results from the base model + DMD2 under the same parameters. SFT generation produced sharper images without the bokeh problem of base + DMD2, but the style was somewhat collapsed: the cinematic stills lacked the 1980s analog film texture, and the last image was not close enough to the 1990s DnD art style.

  12. かみもと · 本地AI视频工作流58

    A complete guide to training an Iris-3B character LoRA on a 12GB GPU with ai-toolkit

    The author trained a character LoRA for Iris-3B with ai-toolkit on an RTX 4070 (12GB); at 512/768 resolution, 2000 steps took about 1.5 hours with about 8.8GB of VRAM usage. The anime character LoRA only learned hairstyle and eye color, with low style similarity, and generating at sizes above the training resolution produced 16px grid-like noise; the photorealistic version of the LoRA performed noticeably better.

  13. Reddit StableDiffusion · 创作者与工作流分享22

    When generating video from reference images with Minimax H3, what to do if a phone or camera always appears in mirror reflections?

    A creator used Minimax H3's reference-image-to-video feature to make an all-POV "Get Ready with Me" influencer-style video. The footage itself looked good, but the person always held a phone or camera in mirror or window reflections, with the problem appearing in about 80% of generations. The creator tried not mentioning the camera, negative prompts, using reference images without a phone or camera, turbo LoRAs such as pmad and dmad, and having the person do other things with their hands, but the phone still appeared and floated around the person; no solution has been found yet.

  14. note · catap_art3d · 图像视频创作分享55

    Testing the MOSS-Image image generation workflow in ComfyUI

    The author tried the MOSS-Image text-to-image model open-sourced by Fudan's OpenMOSS team in the latest ComfyUI templates; the model comes in Base and SFT versions and supports three instruction forms: natural language, tag lists and JSON layouts. Generation takes about 30 seconds on an RTX3060 12GB, and the author also used an AI agent or Open WebUI with a local model to generate layout prompts, and said its Japanese text generation capability is close to Boogu-Image.

  15. Reddit StableDiffusion · 创作者与工作流分享40

    User successfully trains Fizgig character LoRA with MOSS-Image in ComfyUI

    The user successfully trained a Fizgig character LoRA based on MOSS-Image in ComfyUI's 8-step LoRA workflow, saying the result is close to the Krea2 LoRA level it can achieve. The training set was 40 high-quality 1:1 images, about 80% faces and 20% bodies, using LoKR (factor 8), learning rate 5e-5 to 1e-4, 70 epochs, seed 42. The workflow produces an image in about 90 seconds on an RTX 3080.

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

    MOSS-Image tested to accept image input; author releases workflow

    The author tested that MOSS-Image can use image prompts, since its Qwen 3.5 9B text encoder should support multimodal input; the test used the 8-step turbo LoRA. The author says the result is not yet perfect, but it opens the possibility of turning MOSS into an editing model, and has released the workflow link.

  17. Reddit comfyui · 创作与工作流分享20

    Seeking ComfyUI workflow: fix anatomy and shaky line art, reinterpret concepts with Illustrious or Anima

    A user is seeking a ComfyUI workflow to reinterpret their artwork concepts with Illustrious or Anima in order to fix anatomy and brighten the darkest colors. They say results from Gemini, Meta, and GPT have similar anatomy and do not follow the brightening instruction. The user stresses this is a one-off experiment and they have no intention of faking their drawing skill.

  18. Reddit comfyui · 创作与工作流分享22

    ComfyUI user asks for help: hair shows flickering artifacts after upscaling with the latent upscaler in the H3 workflow OBVPM

    A ComfyUI user, after stitching scenes in the H3 workflow OBVPM, upscaled the video with the built-in h3 latent upscaler, and after an 83-minute render the hair turned into flickering "mirage"-like artifacts; they want to know what this kind of problem is called and how to solve it. The user has a 5060 GPU with 16GB of VRAM and asks whether there is a better upscaling method or alternative workflow than OBVPM.

  19. note · kasmintea · 图像视频创作分享15

    How AI illustration creates an autumn air: composition, backlight, and seasonal elements

    The author used AI-generated illustrations to try to add an autumn feel, noting that even when the composition and title are specified to stay level, the image's horizon line still shifts. Backlight and semi-backlight make distant views brighter and foregrounds clearer, and the light and shade of pinecones and the blurring of foreground and background approach the focusing effect of a camera. The author sees seasonal elements as akin to haiku season words: autumn can use red leaves, chestnuts, the moon, and silver grass, while weather and light directly carry emotion.

  20. 宝玉的分享 · 创作与项目分享42

    NotebookLM Video Overview system prompt: the complete creation workflow

    NotebookLM's Video Overview creation skill file has been made public, fully presenting a multi-step workflow from reading notebook sources, helping the user select a topic, and confirming settings to submitting to the video generation pipeline. The prompt focuses on intent disambiguation: distinguishing creating a new video from asking about an existing video, discarding old options when the user restarts, and submitting the user-confirmed topic text verbatim for generation. The file also includes schemas for questionnaires and review cards, formatting and visual style rules, localization requirements, and example tool calls.

  21. Reddit comfyui · 创作与工作流分享71

    How I built an end-to-end AI face-swap video workflow

    The author breaks down an AI face-swap video workflow from raw footage to finished video, using ten folders numbered 00 to 09 to divide stages such as shot planning, reference images, generation packages, frame alignment, final compositing, and quality control. In the generation stage, the pipeline is run twice on ComfyUI with three NVIDIA DGX Sparks, first a 1280×736 draft and then a 1920×1088 refinement; an 8-second 1080p-level clip takes about 1 hour.

    Why it matters: The author reviews the full workflow of a face-swap video with 54 shots and 8163 frames, focusing on easily overlooked engineering steps such as frame alignment and version management.

  22. Hugging Face Daily Papers38

    Any 3D scene in just 1K Tokens: a 3D-native representation for large-scale scene generation

    Researchers propose generating 3D scenes directly in a high-dimensional implicit 3D latent space derived from 2D foundation models such as DINOv2, SigLIP2, and DA3, and using a diffusion Transformer to perform diffusion modeling within that space, avoiding the representation redundancy and limited spatial consistency brought by 2D multi-view or video diffusion. Experiments show that this method outperforms 2D approaches in efficiency and spatial consistency, and validate its scalability as representation and model capacity increase. The paper has been accepted to NeurIPS 2026, and the code has been open-sourced.

  23. Reddit StableDiffusion · 创作者与工作流分享22

    Why is MOSS bad at keeping the same character consistent within a single image?

    A user reports that Qwen 2.1 can stably keep the same character's face and clothing in a single split-screen image, whereas Kromoss based on Moss-base with an 8 steps LoRA loses identity and generates two different people when the character changes pose or scene. The user asks whether this is a prompting issue or whether MOSS itself is not good at same-image consistency.

  24. Reddit comfyui · 创作与工作流分享55

    Author releases MOSS-Image ComfyUI workflow and compares Qwen Image 2.1 with FLUX 3 Klein

    The author released a ComfyUI workflow for MOSS-Image on Civitai, with no payment or login required, and it needs to be used with AusBoss nodes. The workflow can run on a machine with 8GB VRAM and 16GB RAM, produces an image in about 2 seconds at 4 steps on the author's 5090, and includes SeedVR2 upscaling, replacing text in images, frame positioning, and automatic prompting for text-to-image.