Skip to content
Date and source

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

Oct 20Tue
  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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.

  6. 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.

  7. 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.

  8. 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.

  9. 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.

  10. 我爱计算机视觉 · 公开公众号 RSS50

    WorldCast tests 12 visual generation models: can they draw a chair, and can they also "fake" a chair?

    WorldCast built a benchmark covering 8 task categories, 320 handcrafted cases, and 1,352 visual states to be generated, evaluating 12 image models, unified multimodal models, and video generation systems. GPT-Image-2.5 paired with GPT-6 Astra scored 72.8, surpassing the participating video generation systems; when switched to outputting all steps at once as a grid image, the score dropped to 50.1.

  11. Hugging Face Daily Papers32

    WorldStream: streaming generation of interactive 3D worlds with hierarchical planning and persistent state

    WorldStream is an open-ended 3D world model that can extend playable terrain beyond preset maps, respond to open-ended language instructions in a running world, and restore previous edits when revisiting old locations. It uses a hierarchical planner to output structured decision states: Macro fixes the terrain and hydrology structure of a group of tiles, Micro refines it into tile semantics, and deterministic local code then compiles it into terrain, roads, settlements, vegetation and entities. A single world record stores plans, seeds, object identities, mutable states and edits, supporting continuous expansion and interaction under bounded budgets.

  12. Hugging Face Daily Papers40

    StreamEval: interactive streaming video generation model evaluation benchmark released

    StreamEval is a unified benchmark for interactive streaming video generation covering five task categories: virtual human dialogue, virtual human performance, world generation, instruction editing and reference editing. It contains 300 scripted test cases lasting 15-300s, and evaluates on a 0-100 scale across four dimensions: task execution, state consistency, visual fidelity and audio-visual coordination.

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.

  2. Reddit comfyui · 创作与工作流分享62

    LTX-2.5 single-GPU output drops from 130 seconds to 43 seconds: dissecting 384 silent fallbacks per clip in fast attention

    The author did an end-to-end profile of LTX-2.5 (image plus prompt generating 10 seconds of 1080p with audio) on an RTX PRO 6000 Blackwell and rewrote the slow parts; with the same model, same image and prompt, and same 600 W power cap, the original ComfyUI took 130.1 seconds and 19.2 Wh per clip, while the optimized version took 42.9 seconds and 6.9 Wh.

  3. Reddit StableDiffusion · 创作者与工作流分享36

    The Flux-3 prompt enhancer and fine-tuned model also apply to Flux-2

    Flux-3's prompt rewriting model and system prompt generate detailed prompts that work better than handwritten prompts, and can be used for non-simple image editing with Flux-2 9B. The model is deployed via a llama.cpp server on a separate machine and returns results within seconds, faster than ComfyUI's built-in GenerateText method. When Flux-2 won't cooperate with handwritten prompts, the extra wait may be worth it.

  4. BitShovel · 中文AI视频与图像创作案例58

    Krea 3 editing hands-on breakdown: what is preserved, web credits and itemized API costs

    Krea 3 was released on October 15; this article checks the official inputs and outputs, the web usage guide and the API documentation, and sorts out which details are preserved in the two types of editing examples, swapping furniture and swapping backgrounds. The release page says most native 2K edits take under eight seconds, $0.06 is the 2K base item, and the API separately lists $0.02 per 1K reference image and $0.07 per 2K reference image, with high mode adding another $0.02 per round.

  5. note · rosy_dog6649 · 图像视频创作分享22

    What to do when an AI portrait is blurry after upscaling: size checks and high-resolution steps

    If an AI image is blurry after upscaling, first check the source image's pixel dimensions, display position and save method rather than regenerating directly. In Midjourney's current Upscalers guide, a V8.2 square SD image is 1024×1024, and after Subtle or Creative upscaling it is 2048×2048; HD images are already 2048×2048 and currently cannot be upscaled further.

  6. Reddit comfyui · 创作与工作流分享22

    Huge memory usage difference between ComfyUI Desktop and Portable?

    A user generating video with MiniMax H3 on the same 5090 machine with 32GB of RAM found that the ComfyUI Desktop version nearly filled system memory and made the computer stutter; after switching to the Portable version and copying over the same models and workflow, system memory usage dropped to about 45%, VRAM remained at about 31GB, and generation speed was basically the same. The user says the Portable version was far easier to install than expected and asks whether this is a common problem with the Desktop version.

  7. Reddit comfyui · 创作与工作流分享65

    Using Qwen-Image-Edit-2511 crop repair to fix broken hands in a Z-Image Turbo character LoRA

    The author used Z-Image Turbo (8 steps) plus a self-trained character LoRA (0.8) and then SDXL refine to generate a fixed character; most images showing hands had six fingers, fused fingers and abnormal nails.

    Why it matters: The author tested and compared two approaches, full-image editing and hand-crop editing, and provides a reusable crop-repair workflow with specific parameters.

  8. note · miyuu_2026 · 图像视频创作分享38

    ComfyUI hands-on: generation accuracy comparison of Z-Image-Turbo, Flux.2 Klein 9B, Qwen Image 2.1 and Krea.2 Turbo with English and Japanese prompts

    In ComfyUI, Z-Image-Turbo BF16, Flux.2 Klein 9B, Qwen Image 2.1 and Krea.2 Turbo were each used to generate the same photo of a girl's bedroom with English and Japanese prompts, and the author found no significant difference in generation accuracy between the two languages.

  9. Reddit comfyui · 创作与工作流分享22

    Qwen Image 2.1 default workflow's prompt enhancer draws complaints: can't run locally

    A Reddit user tested the prompt enhancer (PE) in Qwen Image 2.1's default workflow and found that raising max_tokens from the default 4096 to 12000 still wasn't enough: with thinking enabled, all tokens are consumed by reasoning, leaving the output empty; with it disabled, it writes 12k-token instructions that get truncated, producing worse results.

  10. Reddit comfyui · 创作与工作流分享36

    Workflow share: retexturing MakeHuman characters with ComfyUI

    A user built two workflows in ComfyUI to retexture MakeHuman character UVs while keeping the texture layout usable for 3D pipelines. The first workflow rearranges the UV layout and processes textures block by block, with descriptions auto-generated by Qwen or entered manually; the second processes by body part in separate channels, upscaling each block to 2160 px before placing it back into a 2048 x 2048 layout, and supports loading previous results for local edits and mixing.

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

    Seeking workflow advice for character replacement in 1-minute-plus long videos on AMD RX 7800 XT while preserving interacted objects

    A user on an AMD RX 7800 XT (16GB VRAM) is looking for a long-video character replacement workflow that, when replacing the protagonist in videos over 1 minute, preserves objects they hold or interact with, such as weapons, tools, and cups, without excessive generation time. In their attempts, Wan Animate 2.2 failed to correctly replace with an anthropomorphic character; Mocha gave better replacement results but with inconsistent fingers and took too long on a 1-minute video; and the Scale 2 standard workflow ignored the reference image and generated new heads and hands from scratch.

  12. note · shio_batapopcorn · 图像视频创作分享22

    Generating a 30-second "holy sword" crafting video with Seedance 2.5 and Wan 3.0 on Pollo.ai

    On Pollo.ai, the same prompt "Please make a holy sword." was tested with Seedance 2.5 and Wan 3.0 respectively, and both generated the complete forging process within 30 seconds. Seedance 2.5 frames the process as a fantasy narrative of a holy sword's birth, while Wan 3.0 is more like a record of a craftsman completing a piece of work. Both focus on hammering hot iron and quenching in water, but skip details such as attaching the hilt and engraving patterns.

  13. Reddit comfyui · 创作与工作流分享10

    Generating a 1-minute video with Wan in ComfyUI: how to keep characters and scenes consistent across multiple clips

    A ComfyUI beginner is doing text-to-video with Wan, starting with 5-second clips as recommended; the duration and quality are good, but they are stuck when trying to stitch them into a complete 1-minute video: after generating a second clip separately with exactly the same character and scene description written in, the visuals differ greatly. They posted on the Reddit comfyui subreddit asking for ways to keep multiple clips consistent.

  14. Reddit comfyui · 创作与工作流分享8

    ComfyUI errors on an old laptop—is there a way to run it online for free?

    A user's laptop has old hardware (i5-8300H, 8GB RAM, Intel UHD Graphics 630 integrated graphics), and after installing ComfyUI, loading image models frequently throws errors, so they want to find a free way to run image generation online. Google Gemini suggested using Google Colab, but they ran into problems there as well.

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

    ComfyUI errors loading GGUF-format Moss q4 model: "Unexpected architecture type: 'moss_image'"

    A user using a GGUF-format Moss q4 file in ComfyUI intermittently encounters the error "Unexpected architecture type in GGUF file: 'moss_image'", with the same workflow sometimes running fine and sometimes always failing. The user has updated ComfyUI and installed the Moss custom node, uses the w4a8 type for the text encoder, and has not yet found the cause of the error.