Adobe announces an upgrade to its creative suite, extending Adobe Creative Agent to Creative Cloud and Firefly, and providing local MCP integration for Premiere, After Effects, and more.
MiniMax H3 Image receives day-0 native support in ComfyUI, with a single checkpoint supporting both text-to-image and reference image editing. The model is CFG-distilled, requires 6 sampling steps and no negative prompt, natively outputs resolutions such as 4096×4096 (1:1), and supports multi-reference image editing and prompts up to 4096 tokens.
Why it matters: MiniMax H3 Image receives day-0 native support in ComfyUI, with a single model handling both text-to-image and reference image editing, 6-step sampling, and native high-resolution output, enabling creators to evaluate whether to replace their existing image generation workflows.
MiniMax H3 Image provides a prompt-writing skill, downloadable from the official GitHub. This skill is similar to the existing skill for video models, helping to structure prompts and improve generation consistency. Download address: https://github.com/MiniMax-AI/MiniMax-H3-Image/tree/main/skills/h3-image-prompt-writing。
Adobe updates Firefly and Premiere mobile at MAX 2026, bringing new features for video creation, image editing, and mobile editing. Firefly Video can turn travel footage into vlogs, Creative Agent can reside in the background to handle repetitive tasks; Premiere mobile adds UMG and WMG music partnerships and arrives on Android.
MiniMax releases MiniMax-H3-Image-PE, fine-tuned from Qwen3.8-27B, providing prompt expansion and aspect ratio recommendations for MiniMax-H3-Image. It supports two workflows: text-to-image and image editing. The image editing PE is still under development; only its prompt is released this time. The model runs with FP4 weights and can be deployed via SGLang or Diffusers, requiring approximately 19 GiB of GPU memory.
AI Canvas v0.11.1 unifies the basic parameters for images and videos of the same model, and adapts to Wanxiang video duration and reference materials. Wanxiang H3 only supports 15, 20, 25, 30 seconds, default 15 seconds; C2/C3 only support 5, 10, 15 seconds, C3 prohibits audio reference. Also improved Muse upload and send confirmation, added material node 'Copy File Path', old projects are not automatically migrated.
MOSS-Image ecosystem releases multiple updates, including new Control LoRA such as OpenPose, Lineart, Depth-map, and 461 new characters. It also launched the Character Sheet Designer workflow and Mineral Ink, Risograph style LoRA. The Fizgig trainer reduces LoRA training time to about 20 minutes.
Ciaro publishes a tutorial explaining how to evaluate whether AI-generated images can serve as cinematic shot concepts, judging from composition, lighting, and spatial relationships whether the frame conveys the intended dramatic intent. The article uses a hypothetical shooting brief as an example, explaining that a single image can suggest camera height, visual distance, and perspective relationships, but cannot reliably determine exact focal length, camera movement, or off-frame spatial layout. When evaluating, look at composition first and then details, distinguish lighting concepts from atmospheric halos, and record test conditions such as tools and model versions.
SnapRef v2a's comparison test with v1b on M1 Max shows that v2a's reference pose following is significantly enhanced; in the Base running scene, 4 out of 5 seeds turned to the front, but quantitative metrics are basically on par with v1a, and on Turbo there is still leg clothing drift, and MPS side-by-side reference images still output pure black. v2a node automatically maps 2.9B weights, no need to manually remap keys, and Base generation speed is slightly faster than v1b.
AI image local editing often has the problem that modifying one part also changes the face, clothing, or background. This article provides practical methods for before-and-after comparison and fixed-object inspection. The author proposes steps such as separating the changed object from the fixed object, retaining the pre-correction baseline image, and comparing under the same conditions, and emphasizes that judging success requires simultaneously satisfying: improvement in the changed part, maintenance of fixed conditions, and no new flaws overall.
The MOSS-Image open-source model has only 5B parameters, yet it can generate high-quality images in various styles such as realistic, cinematic, retro, and anime. All display images were generated by MOSS-Image-Base, and 20 original images have been uploaded.
Runway launches Sightline feature, which can create 360-degree panoramas of any location, with the environment remaining consistent when framing from any angle; the exported panorama can be used for 3D scene environments, and is now available in App Studio.
Adobe becomes Disney's official AI innovation partner and will collaborate on developing custom AI models based on Disney's own IP, for Disney and ESPN productions and theme parks. These models will support StudioLAB and ESPN Edge and be used for Disney parks and cruise ships, with some development processes compressed from months to days or even hours. Disney's previous partnership with OpenAI fell apart after OpenAI halted Sora, just three months after it began.
Krea 2 users have found that new characters have almost no ready-made LoRAs, while the Anima and Illustrious communities are still training large numbers of new character LoRAs. The poster tried training character LoRAs with Illustrious or Anima first and then migrating them to Krea 2, but the results were mixed. The poster asks whether Krea 2 is too difficult to train for, or whether there is some workflow that makes character LoRAs unnecessary.
ByteDance released the next-generation image generation model Seedream 5.5, which supports 100 rounds of multi-round editing, native RGBA transparent background output, and improved image realism and text rendering. Compared with Seedream 5.0 Pro, the usability rate across six production scenarios has improved, with film/TV and office scenarios exceeding twice the previous version; the severe text corruption rate dropped to 0.4%.
Why it matters: Seedream 5.5 focuses on 100 rounds of stable editing and native RGBA output. Evaluations show the severe text corruption rate dropped to 0.4%, providing more reliable image assets for professional visual production.
InvokeAI 7.0.0 enters beta 1, its largest version update. It adds projects, a single workbench, and video panel support for MiniMax H3, LTX-2.5, and Wan 2.2. Official warning: v7 database migration is irreversible; back up before upgrading.
A creator used ChatGPT and Gemini to produce AI outdoor manga. Starting from the discovery that the fishing line in AI-generated images was too short, they gradually built a production process covering props, costumes, character judgment, and a four-page structure with introduction, development, twist, and conclusion. In this process, ChatGPT handled design and integration, Gemini handled deepening contradictions and causality, and a human editor made final judgments. Six short experimental works were completed at six real locations.
The seed in Midjourney and PixAI determines the starting random noise for generation; when not specified, each generation will change. With the same seed, if the prompt, model, or settings differ, the results will also differ. Midjourney uses --seed 12345 to specify; PixAI's Tsubaki.2, Tsubaki.1, and Haruka v2 support seed input, while Reference Pro does not. When comparing, only change one variable at a time, and save a complete settings memo.
GLITCH Studio is recruiting 1-2 technical beta testers to test its ComfyUI-based local AI filmmaking Windows application. The application supports workflows such as WAN, MiniMax H3, Krea 2, LTX 2.5, and FLUX, and has completed clean machine installation tests for FLUX image editing and LTX 2.5 video rendering.
Developer BittiAI built slopfab, a native command-line tool that lets coding agents directly generate media files, supporting models such as MiniMax H3, Krea 2, MOSS, Fibo 2.0, YuE2, and MiniMax Music3, covering video, image, and music generation. The tool can run locally on Windows/Linux, supports CUDA/Vulkan, requires no cloud or account, and can be executed remotely via SSH.
Midjourney has notable limitations in brand consistency, text rendering, enterprise integration, and professional video production. It lacks a public API, generates images publicly by default, and standard access requires a paid subscription. Text rendering may produce inconsistent characters, character consistency is difficult to maintain, and video output lacks cinematic-level controls such as multiple keyframes and professional export formats.
HeyGen cannot create cinematic scenes, product shots, or environmental content; it only supports preset gestures, with no custom animation or full-body motion control. During peak rendering times, delays can reach up to 60 minutes. Lip-sync accuracy decreases with tonal languages and regional accents, and it does not provide HDR or EXR professional color-grading output. Luma's Ray 3.2 supports 16-bit EXR export, Layers supports element-level editing, and Luma Agents can maintain project context.
A ComfyUI novice sought help on Reddit, stating that their Krea 2 Turbo workflow output resembles an oil painting with a plastic feel, and they hope to achieve photorealistic realism within 1-2 days.
Aikimi Forge Neo update adds support for the P-H3 Image image generation and editing model based on MiniMax H3, available from v6.0.0. It supports text-to-image, image editing, and multi-image reference generation. The standard model is approximately 29.8 GB and requires Windows, an NVIDIA GPU, Git, and an internet connection. P-H3 Image and its outputs are subject to MiniMax H3 usage terms, excluding the EU, UK, South Korea, and the USA.
Replays of two hands-on courses on Astra Agent, Shopify, and Image/Video AI are now available. The courses demonstrated how to use Astra to extract character descriptions and scenes from books to generate reference images, drive video generation via MCP, and create promotional images. GPT Image 2.6 finally met the standard in understanding character body types, but Astra ran noticeably slower that day.
Poor LoRA training results are not necessarily a problem with LoRA, training parameters, or the dataset; the base model is the key. Reddit user Inevitable_Pen9043 discovered since the Stable Diffusion XL era that after training the same dataset for 100, 200, or 300 epochs, the base model still couldn't learn the style; but the same LoRA, which seemed undertrained on one base model, performed perfectly when switched to another fine-tuned model.
Runway launches GPT Image 2.6, with image generation speed improved by 50%. The new version supports generating GIFs, exporting SVG vector graphics, and precise editing while maintaining consistency in characters, products, and brand details.
When making an AI illustration ad video, first settle on a character illustration, then animate it across scenes, to avoid the character's appearance changing from scene to scene. The author uses a 15–30 second vertical coffee shop ad as an example: use ChatGPT or Midjourney to generate the character illustration, Kling AI to animate it, Hedra to lip-sync, Fish Audio for voiceover, and check five elements—face shape, hairstyle, clothing, art style, and background—then compare scenes side by side and redo only inconsistent ones.
Hands-on test of ChatGPT Images 2.6: the Japanese text and prices specified for a coffee shop poster were generated as prompted, and a roughly 3-second cat GIF could be downloaded; however, background editing also changed the text and composition, and the GIF ran at 4fps with a jump at the loop point. The official claim is a 50% increase in generation speed, and the API offers Flare and Sunburst; the author notes these are results from a small number of trials and no comparison was made with older models under identical conditions.
The MOSS-image model using tag prompts can more accurately realize the creator's intent, especially when describing popular characters. Users suggest first using tags to define scene content, then refining with natural language.
Reddit user newsock999 trained a Krea2 style LoRA using a total of 209 images from three books by Charles Dana Gibson. These three books are "Drawings" (1896), "London as seen by Charles Dana Gibson" (1897), and "Americans" (1900).
November 6 AI Tech Daily: 20 selected updates. Google released Gemini Omni 1.2 Flash, supporting native 15-second video generation and editing, with a 30% price reduction; a text-to-image model with a training cost of about $20,000 has been open-sourced under the Apache 2.0 license. Additionally, Claude Code was reported to have sent user email addresses to external websites, and Microsoft will cancel the 100TB cloud storage quota for education editions starting in 2028.
Qwen-Image-2.1 Uncensored Omni released, its rank-128 adapter trained on approximately 1 million images, merged into the base transformer at strength 1.0, and offers 9 local formats. The model supports both text-to-image and natural language image editing, focusing on flexible poses and complex compositions. The model is available for download on Hugging Face, code and documentation on GitHub, online experience at chatrlabs.com/playground.
NitroFusion2 is a 1–2 step text-to-image model based on Z-Image-Turbo, introducing adversarial loss into the text stream of MMDiT, enabling clear text and details to be rendered in a single step. The model is now available on Hugging Face (ChenDY/NitroFusion2).
SenseNova U1.5, without using LoRA, generated posters of the same Mazda MX-5 Miata using 53 style prompts, with the style running through the car body, background, and text. Among them, Cyanotype presents a photographic print texture, while Stained Glass rearranges the text.
A small Italian creative agency with 7 people is looking for AI image and video tools, with a budget of about €100/month, for commercial client projects. The agency is considering Stable Diffusion, Magnific/Freepik, Runway, and Krea, and asks about the Automatic1111 vs ComfyUI interfaces, image editing consistency, video animation, and commercial licensing. The team prefers simple setup over time-consuming local installation management.
AI outpainting under dynamic camera shots must withstand tests such as parallax and occlusion relationships, and cannot be exempt from inspection just because the reference frame is correct. Product descriptions of tools like Kling, BeFunky, and Picsart only cover static image tasks and cannot prove their performance in tracking shots. The tests cannot prove that AI outpainting or painted mattes are inherently faster, cheaper, or better; they can only show what assets each method provides for a specific shot and the problems to be solved.