Skip to content
Date and source

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

Oct 16Fri
  1. ArchiGen AI · 图像生成与ComfyUI实战22

    What to test when comparing AI rendering tools: beyond image quality, also check exported alpha, masks, depth and bit depth

    AI rendering tool reviews should check, beyond image quality, what files can be exported after generation: native resolution and bit depth, alpha channel, object or material masks, depth or normals, lighting and reflection passes, and metadata. JPEG has no alpha, no object IDs and no scene depth, making precise post-production edits difficult; traditional renderers such as V-Ray provide EXR channels like Object ID and Z-depth and can support 16-bit or 32-bit floating-point data.

  2. ArchiGen AI · 图像生成与ComfyUI实战34

    Keeping key data in EXR: the color management workflow for architectural rendering in ComfyUI

    A ComfyUI-oriented architectural rendering workflow should keep the scene-linear EXR as the authoritative master and export only one display-ready PNG/TIFF with an explicit view transform and color configuration as AI input. The image denoise value in ComfyUI's official Stable Image Ultra workflow ranges from 0 to 1; lower values stay closer to the input but do not guarantee preservation of HDR range or embedded profiles. Generated results should be composited back onto the master in a bounded way, and highlight clipping, shadow separation and gradient banding should be checked.

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

    Comparison of FLUX 3 Klein, FLUX.2 Klein 9B, Qwen Image 2.1 and Krea2

    A user compared four models—FLUX 3 Klein, FLUX.2 Klein 9B, Qwen Image 2.1 and Krea2—and found that Krea2 and Qwen Image 2.1 clearly lead the FLUX series in human anatomy and realism, with FLUX 3 performing worst on human anatomy. FLUX 3's prompt-following ability is still considered the best, but the author believes its capabilities may be weakened by excessive and unnecessary censorship.

  4. 小众软件 · 创作项目与工具分享12

    VideoProc Converter AI V7.8 anniversary limited-time free giveaway: AI video enhancement, image restoration and 4K/10K upscaling tools

    VideoProc Converter AI has launched a limited-time free license campaign for its V7.8 anniversary; users can enter their email to apply for a license code to activate the software. The free version uses older AI models and supports AI video super-resolution (up to 4x to 4K), image upscaling (2/3/4x, up to 8K/10K), video conversion, compression, downloading and screen recording; some new AI features are unavailable and free updates are not included. The official site also offers a Lifetime version with a one-time payment including lifetime free updates.

  5. note · promptdesign · 图像视频创作分享22

    Gentle hand-drawn style illustration image prompt collection ⑯: 100 in total, copy to ChatGPT or Gemini to use

    This prompt collection includes about 100 "gentle everyday" hand-drawn style illustration prompts, all in directly copyable JSON format, suitable for tools such as ChatGPT and Gemini. In addition to a 1:1 basic version for each theme, it also provides a 9:16 vertical version, a 1280×670 horizontal version, and background-atmosphere adapted versions, covering themes such as family, life, seasons, regions, and interpersonal connections.

  6. note · promptdesign · 图像视频创作分享22

    Anime-style AI beauty image prompt collection ㉑: 100 kinds directly copyable to ChatGPT and Gemini

    This prompt collection includes about 100 "anime-style AI beauty" image generation prompts, covering everyday scenes such as izakayas, stations, cafes, homes, airports, and offices, all provided in JSON format and directly copyable to tools such as ChatGPT and Gemini. In addition to a 1:1 basic version for each theme, it also includes a 9:16 vertical version, a 1280×670 note header image version, and adapted versions for background, lighting, clothing, time of day, and more.

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

    Comparison of basic editing between Flux.3 klein (distilled version) and Qwen image 2.1 (turbo)

    A user compared the image editing speed of Flux.3 klein (distilled version) and Qwen image 2.1 (turbo) on a 4070s: the former takes 10 seconds for 1mp with euler 16 steps, while the latter takes only 3 seconds with euler 8 steps. The test covered prompts such as adding a robot dog, recomposing, changing lighting, converting to black-and-white comic, expression sheet, clothing variants, and figurine transformation.

  8. Hugging Face Daily Papers18

    SCOPE: Semantic boundary credit optimization for text-to-image prompt engineering

    SCOPE is a GRPO-based reinforcement learning objective that lets a prompt rewriting model output semantic partitions while generating enhanced prompts, pools log-probability changes according to semantic spans, and shares loss coefficients for tokens within a span. On Qwen-Image, SD3, and FLUX.1-dev, SCOPE improves by an average of 9.48% on GeneVal and 7.62% on T2I-CompBench compared with the original GRPO, without additional span rewards or image rendering.

  9. Hugging Face Daily Papers24

    RPO: Routing policy optimization for online diffusion reinforcement learning

    To address the high rollout optimization cost caused by perturbed reverse denoising trajectories in diffusion model reinforcement learning, RPO turns attention routing into stochastic exploration actions, samples key blocks according to native attention weights at several denoising levels, and uses a dual-rollout estimator to provide advantage signals. On text-to-image and text-to-video tasks, RPO improves generation quality while training efficiency is nearly 4× higher than Flow-GRPO-type methods, and at inference time it restores dense attention and retains the original deterministic sampler.

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

    Fantasy, abstract, and creative image prompt collection ①: 100 templates that can be copied directly for use in ChatGPT and Gemini

    This prompt collection includes 100 image generation prompts in fantasy, abstract, and creative styles that can be copied directly for use in tools such as ChatGPT and Gemini. Each theme provides three size versions: 1:1, 9:16, and 1280×670, along with two adaptation versions that change the worldview. All are given in JSON format, containing the prompt, negative_prompt, and aspect_ratio fields.

  11. note · fuckdesune · 图像视频创作分享22

    "My Child's 'Daily Life I Absolutely Don't Want Seen'" AI illustration project: using prompts to generate a character's secret moments

    An AI illustration project titled "My Child's 'Daily Life I Absolutely Don't Want Seen'" released a general prompt requiring the generation, based on an original character's reference image and personality settings, of private-life moments the character would not want others to see, in a single 16:9 image. The prompt emphasizes preserving core character features such as face, hairstyle, and physique, and designing secret scenes from personality contrasts—for example, a domineering character secretly practicing lines, or a quiet character speaking baby talk to a plush toy.

  12. note · aiholic0831 · 图像视频创作分享30

    ChatGPT brings AI beauties "into" the real world: 8 prompt techniques to eliminate the "sticker" look

    When using ChatGPT to composite AI-generated people into real street photos, simply attaching a background image still produces a jarring effect such as floating figures and inconsistent perspectives. The author uses a Shinjuku photo they took as the background, adjusts the integration between the figure and the background through prompts, and adds pedestrians in the background and aligns eye-level height to enhance realism. The article also introduces an advanced usage combining a character setting sheet and an outfit setting sheet to bring a fixed character "into" the real world.

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

    Flux 3 Klein performs acceptably on lighting and atmosphere, but details break down on biological subjects

    After hands-on testing of Flux 3 Klein, a user found that the model performs better than most models on atmospheric scenes such as snowstorms, but its details are noticeably chaotic compared with other current models, and proportions are severely distorted when biological subjects are involved.

  14. ComfyUI Wiki · 工作流与开源创作分享69

    Krea 2 Edit LoRA releases its second generation, split into two weights: identity and general

    Conradlocke released the second-generation Krea 2 Edit LoRA on October 16. The biggest change is structural: it is split from a single LoRA into two weights each with its own role: Krea 2 Identity Edit V2 is responsible for transferring the person in a single photo to a new scene, pose, clothing or lighting, while Krea 2 General Edit V2 handles instruction-based edits such as object removal, color changes, relighting, background replacement, object addition, rewriting sign text and stylization.

    Why it matters: Krea 2 Edit LoRA is split into two weights, identity and general, and provides tested parameters for reference image placement and mixing the two LoRAs.

  15. ComfyUI Wiki · 工作流与开源创作分享65

    OpenMOSS releases 5B text-to-image model MOSS-Image, with same-day ComfyUI support

    OpenMOSS released the 5B text-to-image model family MOSS-Image on October 16, including two checkpoints: the pretrained version MOSS-Image-Base and the general-user-oriented MOSS-Image-SFT.

    Why it matters: The 5B text-to-image model focuses on Chinese and English text rendering and JSON layout control, with same-day support from ComfyUI core nodes, which can be used to evaluate layout-oriented image generation workflows.

  16. note · kind_whale5269 · 图像视频创作分享22

    Why is AI image generation always disobedient? A practical guide to changing only the prompt

    AI image generation often suffers from problems such as insufficient image quality, imbalanced full-body proportions, and repeated failures in local modifications, and adding words like "highest quality" and "ultra-high definition" rarely helps. The article points out that the key is not stacking long prompts, but clarifying what to prioritize keeping, what to change, and what must not be touched, and organizes common failure types, prompt improvement methods, templates that can be copied directly, and correction steps to improve completion. The content is not limited to a specific AI service, and results vary depending on the model and features used.

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

    Has Krea 2 gotten worse? Users report declining generation quality after the release of Krea 3 Beta

    Some users report that after the release of Krea 3 Beta, Krea 2's generation quality has noticeably worsened: using the same prompts and the same Moodboard, the consistency between the image and the prompt has decreased, and anime creatures or characters with anime proportions randomly appear without prompting. The user said similar situations occurred before but at a much lower frequency, and the Moodboard used to be the main source of inspiration; they use it through the Krea website because they cannot run it locally due to the lack of a high-performance GPU.

  18. Envato 官方博客42

    Krea 3 AI image editing feature launches on Envato

    The Krea 3 AI image editing feature has launched on Envato, supporting regional editing, compositing of more than six reference images and custom fine-tuning, and can complete high-quality local modifications within seconds without regenerating the entire image. The model is aimed at refining finalized compositions, products or characters, keeping the original subject, labels, lighting and framing unchanged during editing.

  19. note · kasmintea · 图像视频创作分享22

    AI illustration focused on atmosphere: discussing shadows and airiness using the Anima model as an example

    For local image generation models such as Anima, the author shared illustration techniques for creating a sense of "airiness," the core of which lies in distinguishing "shadow" from "shade" and actively specifying the shadow type in the prompt. Anima's character fidelity depends on Danbooru tags, and its backgrounds are somewhat weaker; AI often automatically generates realistic gradient shadows, whereas switching to flat shadows can actually enhance the illustration feel and reduce information density. The author suggests choosing strong or soft shadows according to the atmosphere, and adjusting shadow color and type to change the appeal of the image.

  20. note · ktcrs1107 · 图像视频创作分享12

    Devil's Friday Week 5: 7 characters and unexpected uses

    For "Devil's Friday" Week 5, the author used image generation AI to create 7 characters in one go, including a 78-year-old knight, with names, outfits, weapons, personalities, and backstories all fully defined. It then turned out, from a member conversation, that these characters might only be used as background mobs knocked out by the protagonist with a single punch, so the author decided to seal all 7 to prevent them from exiting the scene 3 seconds after appearing.

  21. Comfy 官方博客62

    Krea 3 launches on ComfyUI via Partner Nodes

    Krea 3 has launched on ComfyUI via Partner Nodes; it is a new image model from Krea aimed at reference-image-driven generation and editing. The model focuses on faster iteration and lower per-image cost while maintaining image quality, and supports uploading multiple reference images that are called by upload order number, for example taking the person from image 1, the jacket from image 2, and the background from image 3.

    Why it matters: Krea 3, which connects reference-image-driven generation and editing to ComfyUI Partner Nodes, lets you learn how to call multiple references by number and feed results back.

  22. Hugging Face Daily Papers37

    MOSS-Image technical report: a 5B open-source DiT achieving long-text rendering and layout control

    MOSS-Image is an open-source image generation system that pretrains a 5B single-stream diffusion transformer (DiT) from scratch, supporting bilingual Chinese-English long-text rendering and precise layout control, and is equipped with an intent-aware agentic prompt enhancement module. On the long-text rendering benchmark LongLongTextBench, it achieves an average text coverage of 96.3%, ranking first among open-weight models; in the end-to-end system comparison on RealLayoutArena, it ranks second. The team will open-source the model, training recipe, and evaluation framework.

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

    Flux 3 Klein performs poorly as an editing model in Comfy

    A user using Flux 3 Klein as an editing model in Comfy found results far worse than F2 Klein, questioning whether it is a model implementation issue or a problem with the model itself. Although the official benchmark shows F3K far surpassing F2K, the user's test prompts included having a man sit on a sofa in a modern minimalist living room and obtaining a right-side view of the scene, and neither produced satisfactory results.

  24. Reddit StableDiffusion · 创作者与工作流分享18

    Krea 3 vs Qwen 2.1 vs ZIT: a random comparison of text-to-image models

    A user on Reddit conducted a random comparison test of three text-to-image models: Krea 3, Qwen 2.1, and ZIT. The user believes Krea 3 is better than the other models at understanding subtle facial expressions such as biting one's lower lip, but it can also generate unsettling images from simple prompts, which stand out even more because of their realism.

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

    Reddit user defends Flux 3 Klein: excellent image quality, but NSFW content is blocked and quality is reduced

    Reddit user Friendly-Fig-6015 posted that Flux 3 Klein is an excellent model with outstanding image quality and called for an end to negative reviews of it. The model's problem is that when NSFW content is detected, it blocks generation and image quality drops significantly; the user says they just need to wait for LoRAs and an unrestricted text encoder.

  26. note · aoao_ai · 图像视频创作分享22

    When generated images have problems, should you fix locally or redraw: my 3 criteria

    A creator who uses AI to make note header images and Instagram comics shares 3 criteria, with Codex assistance, for deciding whether an image should be locally fixed or redrawn: if the problem is not fundamental and only local, fix within a limited scope; if it affects the overall impression such as the face, redraw; if the root cause lies in the script or settings, revise the original instructions. They also found that specifying eye positions with px numbers did not work, and switching to reference images made the fix succeed.

  27. Reddit comfyui · 创作与工作流分享20

    Three film-style prompts under the same seed: Which golden hour atmosphere is more realistic with Z Image Turbo?

    A user tested Z Image Turbo (bf16) with a custom character LoRA at 8 steps, res_multistep / simple, CFG 1, shift 3, and the same seed (710101), changing only the film-style line at the end of the prompt (Cinestill 50D / 35mm / Portra 400) to compare which produces a more realistic golden hour atmosphere. The author also asks if there are any prompt tips they might have missed.