Skip to content
Date and source

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

Oct 6Tue
  1. Theoretically Media blog62

    Two Open-Source Video Models Released: Prism and Kandinsky 6.0; Nano Banana 2.1 Goes Live

    Tencent's Hunyuan team, in collaboration with Fudan University, released the open-source video model Prism, which natively generates 2K video with audio—approximately 10 seconds at 24fps with 48 kHz audio. It only supports image-to-video, requires an 80 GB GPU for 720p, and has no quantized version yet. The code is under the Tencent-Hunyuan GitHub organization, preview weights are on a personal Hugging Face account, and Tencent has not issued an official announcement.

  2. ArchiGen AI · 图像生成与ComfyUI实战22

    Giving Renderers a Normal Bad Input: A Weak-Input Benchmarking Method for AI Renderers

    ArchiGen proposes testing AI renderers using the weakest source images a studio can normally accept, fixing the camera and key design facts, and performing only one initial generation and one controlled revision. The method requires producing five frames: an acceptable source image, default first pass, controlled fix, delivery export, and a difference overlay, scored by the number of changed invariants, first usable output, and fix time. The article cites Rendero's comparison of 11 AI rendering products, noting that Rendero did not actually execute this test.

  3. note · lovely_hebe1274 · 图像视频创作分享22

    How to Request AI Portrait Corrections? Turning 'Just Optimize It a Bit More' into Specific Instructions

    For the problem of wanting to fine-tune AI-generated images but not knowing how to describe it, a correction instruction method suggests first specifying the base image and locking specific locations, then pairing 'what's unsatisfactory' with 'what you want it changed to,' for example requesting 'the character on the right side of the image closes their mouth with a gentle expression, corners of the mouth slightly raised, no big laugh.' The article provides an input template with fields for use case, top priority change, desired effect, parts to preserve, and changes to avoid, and recommends changing only the most important part at a time and comparing the corrected version side by side with the original.

  4. 阡陌的小站 · 本地创作实战70

    Tutorial: deploying Qwen-Image-2.1 and MiniMax H3 locally in ComfyUI with 8GB VRAM

    Qianmo published a tutorial on generating images and videos locally via ComfyUI on an RTX 4070 Laptop with about 8GB VRAM, using the Qwen-Image-2.1 jailbreak GGUF workflow for images and MiniMax H3's 04_i2v_fl2v_4step_sla_lowvram workflow for videos.

    Why it matters: A complete path to running Qwen-Image-2.1 and MiniMax H3 under 8GB VRAM, including model files, directory structure and parameter configuration, reproducible by comparison.

Oct 5Mon
  1. note · hagiri1312 · 图像视频创作分享22

    Fixing the same character with AI portraits: a method for choosing which features to keep in a trial of note's mascot character expressions

    Japanese creator hagiri used the September 29 character design sheet as a baseline to trial 5 expressions and a thumbnail for note's mascot character with AI, by separately specifying "features to keep" and "features that can vary." The prioritized fixed features are a soft face shape, relatively large brown eyes, straight bangs and shoulder-length fluffy brown hair, a white shirt with a light beige-to-pink cardigan, and a soft anime style, while expression, pose, props, and background change with the article scene. The author emphasizes that this was only a small trial, does not guarantee generating the same face in one go, and did not measure time saved.

Oct 4Sun
  1. InvokeAI 官方发布62

    InvokeAI 7.0.0 alpha 1 released: all-new workbench, canvas engine, and video panel

    InvokeAI released 7.0.0 alpha 1, calling it its largest update ever, with about 320 PRs and 2,600 commits, requiring Python 3.12, and with the new interface as default.

    Why it matters: InvokeAI 7 integrates canvas, workflows, and video generation into a single workbench, and adds support for models such as MiniMax H3 and LTX-2.5, which can be used to judge whether to migrate an existing image generation workflow.

  2. note · aoao_ai · 图像视频创作分享50

    AI comic backgrounds always inconsistent? A practical method for fixing room layout with a "background setting sheet"

    When drawing manga with AI, the positions of doors and windows, furniture layouts, and room sizes change every time the same "home" or "nursery" is generated. To address this, the author created a "background settings sheet" independent of the character settings sheet, recording basic layouts such as location type, door/window and large furniture positions, and room relationships, along with composition and description rules—for example, never using a frontal composition for the entryway, fixing the refrigerator position, and fixing equipment like the nursery slide. Small items and items that appear only in a given episode are not included in this sheet and are managed on the script side; the settings sheet is fed directly to the image generation AI rather than being rewritten into prompts each time.

  3. note · lucky_allium8251 · 图像视频创作分享31

    DALL-E 3 Doesn't Read Japanese Directly: It Internally Translates to English First, Then Generates Images

    Independent verification by UX designer Yennie Jun shows that DALL-E 3 does not directly process Japanese prompts; instead, it first internally translates and expands them into English before passing them to the image generation layer. The verification covered 6 languages, with 20 images per prompt for a total of 120 images, and found that Zulu-language prompts generated African-descended characters at a rate exceeding 50%, while English prompts generated them at only about 20%.

  4. ArchiGen AI · 图像生成与ComfyUI实战38

    ComfyUI Architectural Rendering Tutorial: Lock Lighting First, Then Fix Glass Reflections

    The correct order for fixing reflections in ComfyUI architectural rendering is to first lock the final lighting, then use a glass mask combined with depth or Canny guidance to repaint only the glass areas at low denoise. The official ControlNet examples note that the Apply node does not automatically generate depth or Canny maps—users must provide matching inputs themselves. The official inpaint workflow also only repaints the selected area. The author uses five checkpoints for quality control and finally confirms with a difference mask that changes have not exceeded the mask boundaries.

  5. Aitrepreneur · YouTube 创作视频62

    Aitrepreneur Releases a ComfyUI Multi-Image Editing Workflow for Qwen Image 2.1

    Aitrepreneur showcased a custom Qwen Image 2.1 ULTRA ComfyUI workflow supporting multi-image editing with up to 9 images, allowing users to replace characters, remove or move objects via text prompts. The workflow also includes operations such as inpainting, outpainting, and background removal, and is used in conjunction with additional tools and LoRAs.

Oct 3Sat
  1. TechCrunch · AI22

    Pope Leo XIV publicly criticizes AI-generated art

    Pope Leo XIV posted on X calling for protecting human art from AI encroachment, saying there is an ontological difference between art and content generated by machines based on statistical computation over millions of others' images, and that algorithms lack the spark of humanity. In May of this year he detailed his stance on AI in an encyclical, emphasizing human dignity in a world of automation. According to The New York Times, Anthropic lobbied the Vatican to reconsider its position on non-human consciousness, with little effect.

  2. note · ai_to_hibi_note · 图像视频创作分享36

    Writing 8 short stories with local AI on an ordinary laptop and turning them into narrated videos: a hands-on test of Qwen3.5, FLUX.2 klein, and VOICEVOX

    On an ordinary laptop without a GPU (Ryzen 7 7735HS, 32GB RAM), local AI was used to have Qwen3.5 (4B) write stories, FLUX.2 klein 4B draw illustrations, and VOICEVOX provide voiceover, producing a 40.9-second vertical narrated video with 6 scenes.

  3. note · sakisuta · 图像视频创作分享34

    How to choose between GPT Image 2.5's Flare and Sunburst models? A summary of 6 editing features

    OpenAI's GPT Image 2.5 includes two built-in models, Flare and Sunburst: Flare is the default model, with higher image quality than GPT Image 2.0 and 2–4 times the speed; Sunburst is the highest-precision model, slower to generate but with finer control, producing more realistic clothing materials and photographic texture under the same prompt.

  4. ArchiGen AI · 图像生成与ComfyUI实战22

    4K Is Not a Delivery Requirement for Architectural Rendering: Five Steps for ComfyUI Image Enhancement

    ArchiGen argues that "4K" in architectural rendering only specifies file dimensions without clarifying which geometric structures must be preserved or which defects matter. The article recommends writing a one-page output contract before downloading a ComfyUI workflow, specifying exact pixel dimensions and aspect ratio, categorizing areas as protected/editable/discardable, naming specific defects for each improvement, establishing two-tier review standards for delivery size and inspection zoom, and saving aligned originals, outputs, and difference maps as proof.

  5. ArchiGen AI · 图像生成与ComfyUI实战22

    ArchiGen: Tool Selection Comparison Pages Need Traceable Evidence

    ArchiGen proposes replacing generic tool rankings with six documents that allow a second person to reproduce why two tools made the shortlist. These include an identical input brief, input path records, a change log, timing for the second qualifying image, and a real cost table, with costs calculated per qualifying image rather than per generation. The article also notes that Chaos's six-tool guide recommends its own Veras, with comparisons covering monthly, annual, student, pay-as-you-go, and contact-sales pricing.

Oct 2Fri
  1. note · kozo_t · 图像视频创作分享15

    Starting AI image generation in your 60s: A beginner's "I can do this too" experience

    A Japanese author in his 60s shares his journey of starting with ChatGPT and using everyday Japanese to write prompts for AI image generation, saying you don't need to know how to draw to become "a creator." He has actually created images of Showa-era school lunches, calligraphy for longevity celebrations, and manga-style flyers, and has also encountered issues with inaccurate Japanese character forms such as "ju" (longevity) and "shuku" (celebration), requiring repeated corrections. He summarizes five beginner tips: describe characters specifically, include locations, explain the atmosphere, specify the style, and don't hesitate to revise repeatedly.

  2. あいきみ · 图像编辑与LoRA实测60

    FLUX 3 Image hands-on test: Retention and boundaries across 10 consecutive edits, 10 reference images, and 4K generation

    The author conducted a 10-consecutive-edit test on FLUX 3 Image (BFL API), using each output as the next input and verifying retention through same-coordinate RGB pixel comparison. A fixed 10,720-pixel facial region remained 100% identical to the original image after the first 5 edits, dropping to 52.67% by the 10th edit; after the 6th edit expanded the edit range, facial lines and shadows outside the specified area also changed.

  3. Envato 官方博客40

    FLUX 3 Image launches in Envato AI image tools

    Black Forest Labs' FLUX 3 Image has joined the Envato AI model lineup, supporting multi-reference image generation and editing, native 2K/4K high-resolution output, in-image text generation and modification, reference-based style transfer, and precise local editing. Users can specify the exact role of each reference image and add, remove or replace specific elements while keeping the rest of the image unchanged.

  4. Midjourney 官方更新22

    Midjourney Alpha update: style previews, default parameters and a collaboration tool teaser

    Midjourney launched sidebar style previews, larger Create images and folder default parameter settings for alpha.midjourney.com; this week focused mainly on bug fixes and cleanup. Multiple srefs in the prompt bar are now merged into a single style pill with a thumbnail, hovering shows the code, and --exp can be pinned as a folder default. The company said new collaboration tools will launch next week and teased persistent edit history and better ways to explore styles.

  5. ArchiGen AI · 图像生成与ComfyUI实战22

    Veras material generator: reference images have color, but no scale

    Chaos says Veras's material generator can turn any reference image into a material, or match a material in the Cosmos library, while preserving mapping, scale and alignment. But the article points out that an unmeasured material sample is only a color reference, not a dimensional specification, and generated results may look similar while corresponding to different products. It recommends placing a ruler next to the sample before uploading, checking maps channel by channel, and verifying scale and repetition under three lighting conditions: diffuse, grazing and indoor light.

Oct 1Thu
  1. note · uuhai0625 · 图像视频创作分享22

    An office worker who made 10 LINE sticker characters with image-generation AI sorts out what it is really good and bad at

    An office worker who made 10 LINE sticker characters with image-generation AI sorted out the strengths and weaknesses of image-generation AI. It is good at quickly producing diverse design options, expressing atmosphere and texture, and faithfully following composition and lighting instructions; depicting hands and fingers remains a weak point in 2026, while text rendering and character consistency have clearly improved, with progress in areas such as Midjourney v7.

  2. note · yaoyoroztech · 图像视频创作分享30

    Two image AIs released the same week, and one's images can't be posted — Qwen-Image 2.1 and Ming-Image 0.1 differ not in performance but in license

    Two image AIs released the same week, Qwen-Image 2.1 and Ming-Image 0.1, differ not in performance but in license. Qwen-Image 2.1 was released on September 20, its drawing component is 7B, it supports transparent backgrounds and up to 10 reference images, and ComfyUI supported it on the first day of release, but it uses a research license, with commercial use requiring a separate agreement, so the author decided not to publish its generated images.

  3. note · shiori_ai_diary · 图像视频创作分享22

    AI Image Generation Is So Fun That the Output Folder Completely Lost Control: Developer Casually Built MediaRanker and Listed It on the Microsoft Store

    Because Asakura Shiori had too many AI-generated images to choose from, the developer created a Windows app called MediaRanker that lets users browse images while rating them with ★ marks, and it has been listed on the Microsoft Store. The app supports continuous browsing of images and videos within a folder with direct marking, and the Free version is free with no time limit. The developer stated that more features will continue to be added.

  4. note · glad_snake1405 · 图像视频创作分享38

    The Golden Rule for 'Pose Specification' in AI Image Generation: Using the 3-Point Anchoring Method to Get Midjourney and Stable Diffusion to Produce Images as Expected

    To address issues of pose loss of control and distorted limbs in AI image generation, the '3-Point Anchoring Method' can be used to break down prompts: gaze direction, joint (skeletal) position, and contact with objects or the ground, replacing abstract terms like 'cute pose' and 'dynamic pose.'

  5. ArchiGen AI · 图像生成与ComfyUI实战22

    Why AI Rendering Rankings Change After Super-Resolution

    ArchiGen points out that current AI rendering comparisons often mix three different outputs—native generation, vendor super-resolution, and manual post-processing—into a single score, rendering image quality scores meaningless. The article recommends preserving unprocessed native files and recording pixel dimensions, format, generation mode, and seed, while separately annotating tools and parameters for each stage: generation, denoising, super-resolution, repair, color grading, and export. Native fidelity and delivery readiness should be tested separately with two scores published; if the number of protected details changes after super-resolution, fidelity is deemed non-compliant.

  6. Theoretically Media blog62

    Early Hands-On with Kling 4, Plus a Look at Gemini 4 Argon, Ideogram 4.5, and Eleven v4

    The author conducted hands-on testing of the Kling 4 early access version, which currently only offers the Flash model with 720p output, available to Ultra subscribers on annual plans; larger models are planned for October. Testing shows that Kling 4 appears to have resolved the lip-sync drift issue from 3.0, supports up to ten keyframe references, but performances are somewhat flat.

  7. Luma News62

    Luma Launches Variants, Automatically Adapting a Single Ad to Multiple Sizes and Languages

    Luma introduces Variants: after a user uploads an approved static ad, they select a layout and language, and the system automatically generates versions for each size and language combination, keeping the logo, headline, and CTA unchanged while only adjusting layout and copy.

    Why it matters: Ad teams can use this to evaluate batch-adapting a single approved asset into multiple sizes and language versions, replacing manual rework.

  8. ArchiGen AI · 图像生成与ComfyUI实战22

    Stop Giving Photorealism a Single Vague Score: ArchiGen Proposes a Six-Category Defect Checklist for Rendering Quality Control

    ArchiGen proposes replacing the vague 'photorealism' score with a 'defect checklist,' breaking rendering quality control into six categories: geometry, material, lighting, reflection, scene context, and image completeness. Each item requires marking the location, describing the deviation, and stating whether it impedes the intended use. Geometry has first veto power; if geometry is wrong, one should return to the control stage to adjust inputs, masks, edges, or depth guidance and generation strength, rather than handing it off to super-resolution tools.

  9. Luma60

    Luma Integrates Ideogram 4.5 Image Editing Model

    Luma announced that Ideogram 4.5 is now available on Luma, and users can use it at app.lumalabs.ai, giving creative teams more control over deciding which content in the image changes and which is preserved. The cited official Ideogram introduction states that Ideogram 4.5 is a model focused on precise editing that can eliminate artifact accumulation in multi-round editing, thereby supporting multi-round editing; it is already available on Ideogram, the API, and launch partners, with open weights coming soon.