Skip to content
Date and source

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

Oct 20Tue
  1. Kapwing 创作指南69

    How to build an AI film workflow with GPT-6 Astra, Blender, and Seedance 2.5

    Kapwing released a tutorial demonstrating using GPT-6 Astra to connect with Blender to build a dragon and scene's gray model, animation and camera shots, then using GPT Image 2 to generate design drawings and a photorealistic first frame, and finally Seedance 2.5 to generate a 10-second continuous shot.

    Why it matters: The author tested using GPT-6 Astra to drive Blender to build a 3D scene, then went through image-to-3D and Seedance 2.5 to produce the video, and listed a comparison of three 3D models and itemized costs.

  2. Kapwing 创作指南73

    How to use Seedance 2.5 to turn a single piece of footage into 4 camera angles

    Kapwing released a tutorial introducing the use of Claude skills to generate camera instruction prompts, then using Seedance 2.5 in Kai to reshoot a single fixed-camera piece of footage into a multi-camera edit.

    Why it matters: The author tested single-camera to multi-camera conversion with Seedance 2.5, providing the complete prompt and failure cases, which can be transferred to multi-camera editing for short dramas.

  3. Stable Diffusion Art 教程78

    MiniMax H3 reference-to-video (ref2va) ComfyUI workflow tutorial

    Stable Diffusion Art released a complete ComfyUI workflow tutorial for MiniMax H3 reference-to-video (ref2va), which can combine up to 9 reference images, 3 reference videos and 3 audio clips to generate a 24fps multi-shot video with synchronized stereo dialogue, sound effects and music in one pass.

    Why it matters: The tutorial provides the complete node configuration and prompt structure for MiniMax H3 reference-to-video in ComfyUI, which can be transferred to multi-shot character consistency workflows.

  4. Wireflow · AI短剧与视频制作教程65

    Field test of 8 Veo 3 prompts on Veo 3.1: 5 passed, 1 partially passed, 2 failed

    Wireflow ran each of 8 Veo 3 prompts once with Veo 3.1 Fast, at 8 seconds, 720p, with sound; the results were 5 passes, 1 partial pass, and 2 failures. The passes were mostly single shots with a single action plus one line of dialogue under 15 words; the failures centered on requiring three shot changes within one 8-second clip, and a 39-word line being truncated to 31 words.

    Why it matters: The author tested 8 Veo 3.1 Fast prompts with the same set of parameters, giving the reason for success or failure and a rewrite for each, which can be compared against your own video production workflow.

  5. かみもと · 本地AI视频工作流58

    Author trains a style LoRA "SepiaBlue Luminous" for MOSS-Image and releases it publicly on Civitai

    The author trained a style LoRA "SepiaBlue Luminous" for the local image generation model MOSS-Image and released it publicly on Civitai, with the trigger word qsv2style. Training used 100 ChatGPT-generated images sharing only the style, with different characters and scenes, based on MOSS-Image-Base, at 1024 resolution, rank 16, learning rate 1e-4, for 1000 steps, and the 400-step version was ultimately chosen.

  6. Reddit comfyui · 创作与工作流分享15

    Help Requested for Error on Second Generation with MiniMax H3 Video Extender (backported)

    The user is using the MiniMax H3 video extender (backported) for video extension; after the first generation, during upscaling in the second generation, the sample custom advanced node inside the model's latent upscale subgraph reports an error. The poster asks about the cause and hopes for suggestions on alternative nodes.

  7. Reddit comfyui · 创作与工作流分享12

    How to Run a Wan Image-to-Video LoRA Workflow Locally with ComfyUI

    A ComfyUI desktop newcomer wants to load a self-downloaded LoRA locally in a Wan image-to-video workflow, but after replacing the LoRA parameters the results are abnormal, and manually adjusting after unpacking the subgraph reports "not enough memory", while it runs normally at low resolution in the packed state. Their goal is a simple IMAGE → PROMPT → RESULTS flow that reproduces the model page's example results.

  8. Reddit comfyui · 创作与工作流分享22

    Too Many ComfyUI Custom Nodes Slow Startup—Any Cleanup Suggestions?

    A user reports that ComfyUI startup has slowed due to installing a large number of custom nodes over the past few months and wants to know which nodes can be deleted. They found a cleanup solution that looks promising and ask whether anyone has used it, whether it works, and whether there is a faster and simpler way, or simply removing everything and reinstalling one by one based on workflow errors.

  9. Reddit comfyui · 创作与工作流分享72

    Evangelion Fan Test: Frame-by-Frame Alignment and Compositing of AI Characters and CG Mecha

    The author released an Evangelion fan concept proof of concept, placing an AI-rendered Sachiel and a Redshift-rendered CG Unit-01 in the same shot to achieve physical contact, with the contact point maintained 1:1 frame by frame during complex camera movement, with no drift.

    Why it matters: The author uses live-action-grade compositing of AI characters and CG characters aligned frame by frame to demonstrate a concrete approach for integrating AI layers and depth channels into a traditional VFX pipeline.

  10. Reddit comfyui · 创作与工作流分享22

    Using Flux 2 Klein for i2i face swap: how to make expression, lighting, and position match exactly?

    A user asked for help in the ComfyUI community: when doing i2i face swap with a Flux 2 Klein workflow, the generated result does not match the original image in head size, position, and facial expression, and the eyes, mouth, and eyebrows cannot be precisely aligned. The user had previously first trained a model to generate "good enough" photos, then did the face swap in the final step, but the result was still unsatisfactory, and asked whether this goal is still unachievable at present.

  11. Reddit comfyui · 创作与工作流分享28

    The conflict between Qwen 2.1's transparent background capability and the prompt enhancement template

    Qwen 2.1 is well received for being able to generate images with transparent backgrounds, but the default model used for prompt enhancement in its "official" template, qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot, and all its variants, carry explicit instructions that "silently" change transparent background requests into pure white backgrounds. Users can bypass this by "jailbreaking" with forced transparency instructions or writing their own prompts, but it is not 100% reliable; even with an explicit transparency request, a pure white background is still often produced.

  12. Reddit comfyui · 创作与工作流分享18

    MOSS 5B model architecture diagram released: drawn based on official code, modular implementation using Typst and CeTZ

    For the MOSS model released 4 days ago, someone drew a MOSS 5B architecture diagram based on the official code, along with supplementary information from the paper and code review. The diagram is drawn with Typst's CeTZ, uses a modular design, can render each module separately or be reused in other diagrams, and is open-sourced under the MIT license.

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

    New AI beauty image prompt collection ⑯: 100 JSON prompts that can be copied directly into ChatGPT and Gemini

    A collection of AI beauty image prompts for tools such as ChatGPT and Gemini, about 100 in total, provided in JSON format and ready to copy and use. In addition to the basic version, each theme also includes 9:16, 1280×670, background adaptation, and atmosphere adaptation versions, covering urban scenes such as an old Madrid cinema and reading on the New York subway. The prompts contain prompt, negative_prompt, and aspect_ratio fields, aimed at uses such as note header images, SNS, and short video materials.

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

    Say goodbye to stiff standing poses! Learning action lines and center of gravity with ChatGPT | ponko's AI image creation training ①

    Japanese creator ponko uses ChatGPT as a teacher to learn to design action lines and center of gravity before AI image generation, in order to escape stiff character standing poses. The first assignment was the pose of "jumping away in surprise at an explosion, at the moment of landing"; she first broke it down herself as the center of gravity falling on the bent right leg, the right hand supporting on the ground, and the left hand guarding the body, then had ChatGPT critique it. She found that each limb has its own role and that camera position changes the readability of the action, and that vague instructions like "feels good" cannot convey intent.

  15. note · kasmintea · 图像视频创作分享18

    How AI illustrations create a sense of air: a composition review

    The author used the same prompt to have AI generate an illustration thumbnail, asking for foreground blur, a background with depth, and ground to enhance the sense of space. The result showed the character's clothing as requested, but the title text was unexpectedly tilted, the top and bottom of the image were cropped, and the composition had confused visual guidance. The author believes the sense of air itself (backlighting, blurred street reflections, petals and hair) is fine; the problem is the composition: the title rises toward the right shoulder while the hand on the railing extends toward the back left, making the whole uncoordinated.

  16. note · imamirai_editors · 图像视频创作分享30

    ChatGPT image generation: Aizawa freely releases prompts for five scenes: morning, noon, evening, night, and late night

    Aizawa posted ChatGPT image generation prompts for five scenes—morning, noon, evening, night, and late night—on X, free to use. The prompts require attaching a reference image of an adult character aged 21 or over that the user has the right to use, to maintain face shape, hairstyle, hair color, and eye color, and uniformly adopt a delicate, clean anime illustration style with soft light and continuous shadows. The five scenes correspond respectively to morning holding a cup with both hands, afternoon showing a blue-purple color swatch, evening holding a cup indoors in full body, night resting on a sofa, and late night by the window before turning off the lights.

  17. note · seisei_surumono · 图像视频创作分享20

    ComfyUI virtual human node development, part four

    ComfyUI virtual human node development update: after improving the character's personality, replies became longer, and the increased data brought a feeling of "intellectual improvement." Added background color and image switching, environment recognition capability, and expanded monologue and humming vocabulary (previously only words like starry sky and fluffy), and also registered pronunciation corrections in the VOICEVOX dictionary.

  18. note · sasao_66 · 图像视频创作分享22

    Research on DomoAI fireworks festival date video generation prompts

    Using DomoAI's AI video feature, a fireworks festival date clip was generated from one image plus an English prompt: boyfriend first-person POV, a woman in a yukata turns around and says "きれいだね!" and holds hands, with fireworks blooming behind her. The prompt also required a natural Japanese female voice and accurate lip-sync; the author rated the result as "beautiful but a bit funny" and said performance was good. DomoAI free generation includes a watermark, and the author recommends subscribing.

  19. note · aiholic0831 · 图像视频创作分享38

    Generating a Weekly Shonen magazine cover with a one-line ChatGPT prompt: from default cover to custom magazine name and serializations

    The author used ChatGPT to generate a Weekly Shonen magazine-style cover in 3:4 ratio; a one-line prompt automatically fills in the magazine name, issue number, photo subject, and serialization titles. After writing the magazine name, issue number, price, and title text all into the prompt and restricting it to "use only these words," a custom cover of 《週刊少年アイホリ》 2026 No.46, November 4 issue was obtained, with all specified text rendered correctly. The author also found that swimsuit-type content is easily rejected, and that a large waist-up figure gives a stronger cover feel.

  20. ArchiGen AI · 图像生成与ComfyUI实战40

    How Veras Smart Selection differs from Cryptomatte: semantic selection is not scene identity

    Veras 4.5's Smart Selection can quickly select visible pixels by category, object, material, or custom prompt; the selection map is cached and restored when switching renders, making it suitable for quick local edits. However, it only infers regions from the image and does not inherit object names or material slots from Revit, Rhino, SketchUp, or the renderer; Cryptomatte, by contrast, records identities assigned in advance by the scene and is suitable for repeatable compositing. In May 2026, the official side said color segmentation map export was not yet supported, and the team plans to add it later.

  21. ArchiGen AI · 图像生成与ComfyUI实战38

    Veras 4.5's Describe Source Image can describe an image, but architects cannot use it directly as a prompt

    Veras 4.5 moved Describe Source Image into the main panel, where Florence 2 generates short, medium, and long descriptions that can be copied as a prompt starting point; switching the source image clears the old description. This feature only reports what is visible in the image and cannot recognize design intent or fixed facts that must be preserved; copying it directly turns omissions into design decisions. The correct approach is to annotate each sentence as "observed/inferred/missing/misnamed," then split fixed geometry and requested changes into two sections in the prompt.

  22. TelosCompass · note AI视频创作实践38

    Making a person in a photo speak with DomoAI: check in the order of voice, lip-sync, face

    DomoAI's "AIアバター" feature can make a person in a photo speak: put the dialogue in the audio field, and write only actions and expressions in the prompt field—do not copy the dialogue. The selected seconds do not necessarily equal the final video length; after generation, compare playback time with the recording length to confirm the last line is complete. Check the final video in the order of voice, lip-sync, face, and if there is a problem, change only one item at a time and redo.

  23. note · noble_mink9287 · 图像视频创作分享12

    A university student's side-hustle practice of making AI videos and posting them to TikTok: making and spreading are two different things

    A university student made AI videos on the theme of "Future Tokyo" and posted them to TikTok, discovering that making a video and getting people to watch it are two different things, and that AI videos are not finished once made. During production, images and videos had to be adjusted repeatedly, with more trial and error than expected; only after posting did they realize they needed to think about what kind of content sparks interest and how to make people watch to the end.

  24. note · nobinlog · 图像视频创作分享38

    Detailed explanation of ComfyUI-MangaScreen settings and how to improve manga-style illustration quality with Anima

    ComfyUI-MangaScreen is a tool in ComfyUI that can generate relatively high-quality manga screentone; the author tested the effects and caveats of settings such as screen_lines, white_point, black_point, light_tone, dark_tone, gradation, pen_threshold, and white_lines one by one.

  25. note · kongo_jun · 图像视频创作分享40

    Hands-on test of MOSS-Image's 8/4/2-step acceleration LoRA in ComfyUI

    Hands-on test in ComfyUI of MOSS-Image sft's 8/4/2-step acceleration LoRA: on an RTX 4090, 8-step generation took 10 seconds, 4-step 9.5 seconds, and 2-step 9 seconds, all with 16GB VRAM. After the model loads, a single image can be produced in about 1-2 seconds, and reducing steps still generates some text. The LoRA is relatively large and requires about 1GB of extra VRAM.

  26. note · ai_nemui · 图像视频创作分享30

    MOSSimage local image generation hands-on test: INT8 load close to SDXL, LoRA training failed

    A hands-on test of MOSSimage local image generation showed that at INT8 precision the load is not much different from SDXL, while BF16 is close to the heavier version of anima and clearly lighter than Qwen. The author used standard tests such as reproducing the 《Seiken Densetsu 2》 cover, and believes that models able to draw this kind of fantasy background locally were previously rare, and that Krea2 could not do it. LoRA training failed due to overfitting, and in a test of locating the tail end of manga speech bubbles, the tail ends generally deviated from the mouth by tens to hundreds of px.

  27. Loomance · 本地AI动画实测67

    Hands-on test of Wan 2.2's VRAM and RAM usage on an 8GB graphics card

    The author tested Wan 2.2 image-to-video on an RTX 5070 12GB; with roughly 34GB of model weights, the GPU actually uses only 8–9GB of VRAM, while the rest is dynamically loaded by ComfyUI and streamed from RAM and disk.

    Why it matters: The author reran the test on the same machine with RAM limited to 16GB, correcting his earlier conclusion that treated RAM as the bottleneck, and provided reproducible measured data on VRAM and RAM usage.

Oct 19Mon
  1. Reddit comfyui · 创作与工作流分享22

    How to make two character LoRAs interact naturally in the same Illustrious image in ComfyUI

    A ComfyUI beginner trained two character LoRAs, one male and one female, on Civitai, paired with the WAI-Illustrious-SDXL checkpoint. Each works well when generated alone, but when generated in the same image there is feature mixing, character duplication (2 women, 2 men or even 3 people) and reduced fidelity.