Guizang shares Grok Bot studio configuration
My Grok Bot studio configuration
My Grok Bot studio configuration
This tutorial explains how to set up an environment on Windows using the ComfyUI integrated package from Bilibili creator 秋葉aaaki; after unzipping, double-click the one-click launcher to open the interface. Built-in templates can be used directly, but models must be downloaded separately. The author uses WorkBuddy to import the integrated package folder and completes the initial configuration through conversation, downloading the qwen-image-2.1-turbo workflow and setting up text-to-image, image-to-image, and image-to-video, among others; WorkBuddy can also automatically build and execute workflows via API.
A quick guide on how to get Grok Bot to consume WeChat official account articles here: https://mp.weixin.qq.com/s/1afnn-WeU4KvD03lqSIv2w
The theme song MV for the AI animated film "The Owl of the Wind Forest" has been completed and released first on YouTube. The theme song was made with Suno AI, the still images were generated with Nano Banana in Google Flow, and the main film uses Seedance2.0 and Seedance2.5 for video generation.
Aimed at the need to bring AI image generation in-house into one's own production environment, the article outlines the division of labor between the two routes of Stable Diffusion and FLUX: the Stable Diffusion family is suited to running on local or cloud GPUs.
The ArchiGen article argues that when a project's relationship with the city is itself a design proposition, the urban context should be built as a navigable model rather than a static backdrop. It cites a Chaos-sponsored Architect Magazine case: Morevi Studio's 84-hectare proposal for Ambassadori Island Batumi in Georgia builds Batumi and the artificial island as two linked zones placed in the same scene, enabling real-time review of cameras and atmosphere.
AI video generation models are neural networks that learn from massive amounts of video clips and generate every frame from scratch; given text or an image as input, they output short videos of 2 to 30 seconds at 360p to 4K. Mainstream architectures fall into two categories: diffusion models, which progressively denoise to generate the entire video, and autoregressive models, which predict subsequent content block by block. The VIBE app includes 39 video models such as Kling, Sora, and Veo; the version number determines duration, resolution, and whether audio is included, while suffixes such as Fast, Pro, and Mini correspond to trade-offs in speed, image quality, and size.
OpenAI Codex can use CoAnimator to edit video projects and call the renderer to output MP4: Codex is responsible for reading requirements, planning scenes, writing project files, and making revisions based on feedback, while CoAnimator handles preview and rendering.
Wan 2.2 image-to-video can run on an RTX 5070 with 12GB of VRAM; a 57-frame clip takes 10 minutes 38 seconds at 832×480 and about 67 minutes at 1536×864, both including frame interpolation and upscaling to 2048×1152.
AITuber has released a step-by-step guide to making AI rap videos: first prepare the audio track, either by uploading MP3, WAV, M4A, AAC, or OGG (max 50MB, 10 minutes), or by using the built-in song generator to generate from a description or custom lyrics (up to 5,000 characters), which returns two versions each time. Then select a single character and art style to generate the visuals, align subtitles to each word, and export a 9:16 vertical video; a length of 30 to 60 seconds is recommended.
The author used Opus 5.5 with three.js to basically produce a rainy-night convenience store miniature scene in one round, with only clipping fixed in the second round, all within the 5 hours limit of the cheapest membership. Prompt highlights include a third-person view miniature 3D model that can be dragged to rotate and zoom, a Japanese-style convenience store street corner as the main subject, a warm-cool contrast between the bright warm interior and the wet rainy night outside, cel-shaded cartoon rendering, and animated touches such as continuous rainfall and puddle ripples.
The author trained a Qwen Image 2.1 character LoRA with ai-toolkit 0.13.19 on Google Colab's A100 80GB; 1,000 steps took about 21 minutes this run, whereas the same number of steps on a local RTX 4070 12GB takes about 2 hours.
Stable Diffusion Art publishes a tutorial on how to build a MiniMax H3 text-to-video workflow in ComfyUI, generating a 5-second 24fps clip with stereo sound in one pass.
Aitrepreneur releases a tutorial demonstrating how to train a LoRA for Qwen Image 2.1 EDIT with AI TOOLKIT locally or on RunPod, using under 10GB of VRAM. The video also covers building your own training dataset, setting training parameters, and comparing the trained LoRA's results in ComfyUI.
Using DomoAI's "text-to-video generation" feature, write the prompt around four items: protagonist, location and lighting, action, and presentation style. For the first video, do not rewrite with "AI optimization"; generate directly, then compare the result against the prompt word by word to sort out which words were rendered and which were not. Save the rewrite feature for the second video; Runway has a similar feature. Store the comparison results along with the prompt, aspect ratio, duration and points to change in the same folder as the video, and add only one unimplemented word each time.
Google AI Studio has notified some developers to switch the Gemini API from postpaid to prepaid by October 12, after which service will be interrupted. When the prepaid balance is zero, paid Gemini API calls will fail even if the API key and code are fine; the balance is managed by Cloud Billing account rather than by project or API key, and is valid for 1 year after purchase. Users who only use the free tier do not need to migrate.
The author tests MiniMax H3's Multiframe Reference in ComfyUI on an RTX 5060 Ti 16GB to verify whether it can use still images as storyboards.
On October 11, "Blink Day," author mii♪ shared a set of free prompts: first use ChatGPT to turn pet photos into 9:16 vertical watercolor picture-book illustrations, then use Google Flow or Gemini to convert the illustrations into roughly 8-second videos, so that when the pet winks one eye, the lantern and magic symbols flash in sync. The prompts support five themes—star, heart, moon, flower, and four-leaf clover—with customizable symbols; the finished works can be used personally or posted to SNS, and reposting and secondary distribution are prohibited.
AI filmmaker Frank Houbre disclosed his 2026 creative tool stack, divided into six stages according to the filmmaking process, with only one main tool kept for each key stage. The writing stage centers on ScreenWeaver, which he co-created, with Claude and Claude Code handling review and repetitive tasks; image and video generation center on Higgsfield as the main tool because it aggregates multiple models under the same account.
A composition tutorial for generative AI illustration creators, explaining how to create heart-fluttering scenes in the style of romance manga using male and female characters. Core techniques include low-angle upward shots to create the tension of "being looked down at," high-angle downward shots to emphasize cuteness and protectiveness, rule-of-thirds offset composition to create negative space and narrative, and specific composition types such as over-the-shoulder glances back and subjective viewpoints. The article includes English prompts that can be copied and used directly, and recommends side lighting and backlighting to enhance three-dimensionality and emotional atmosphere.
The author compiled 15 sets of prompts for generating hyper-realistic photos with Nano Banana 2.1, categorized by camera type: iPhone, 0.5x ultra-wide, smart glasses, gimbal, drone, video call screenshot, GoPro, dashcam, surveillance, compact camera, flip phone, disposable camera, 1990s developed photo, Polaroid, and film SLR, with each set specifying details such as shooting angle, lighting, noise, and compression artifacts.
Midjourney can generate 3D game character design sheets with 5 sets of English prompts; each set outputs four full-body views—front, three-quarter front, side, and back—in the top row, and 6 expression variations in the bottom row, with unified character geometry and clothing. The prompts cover five styles: everyday game style, Eastern fantasy stargazer, 3D animated film witch, realistic sci-fi mecha warrior, and classic RPG Japanese female swordsman, all with Japanese translations and parameters (--ar 16:9 --v 7 --raw --s 100).
An AI cat video YouTube channel passed review again after having its monetization stopped twice and about half a year of zero income. The author reviewed their changes to video content and story, subtitles, sound effects, and BGM editing, use of multiple shots and still images, titles, descriptions, thumbnails, channel description, and the limited-public videos used for re-review, and emphasized that these practices do not guarantee passing the review.
MiniMax H3 RefMod was used in animation production, and in testing the generation quality has reached a level where it can be used fully as animation; both live-action and animated versions can be generated. Making about 55 minutes of video with Comfy-Cloud costs about $80USD per month; enterprise-level costs are about $630USD/month, and operation only requires training an operator. Compared with Google Colaboratory disconnecting every 30 minutes, Comfy-Cloud or the ComfyUI Desktop version is more worry-free and easier to use.
The author introduces RefMod, a new character-fixing technique for MiniMax H3 in ComfyUI; without LoRA training, using 8 to 20 images of the same character, it can generate a 1 to 2MB .safetensors character file within tens of seconds.
Why it matters: RefMod encodes 8 to 20 images into a 1 to 2MB character file, offering creators who do not want to run LoRA training a comparable alternative path.
The author used Wan 2.2 on Hugging Face, invoked via Claude Code's MCP, to hand large movements in 10 still-image storyboard frames to a video generation AI, while small movements were done by moving parts with a Python program. After inputting both the first and last frames, the backgrounds and lines in the videos generated by Wan 2.2 did not break down, and there was no case of the art style being overwritten like with Kling previously; however, the character still stands up when climbing rocks, the leaf-licking only repeats once, and text instructions only fixed the feet, water bottle, and eyes.
An experiment on image generation with Gemini and ChatGPT found that generating a pose reference image with Gemini and then handing it to ChatGPT to generate an illustration makes it easier to approach complex aerial dance poses such as Aerial Walkover, but the generated results still show a left-right reversal problem where the character's left arm is attached to a right-hand shape. The author believes reference images are only pose "clues" rather than the "correct answer," and that after generating an image one still needs to check the left-right orientation of limbs and joint connections one by one; this conclusion comes from personal trial observations, not a statistical comparison of success rates.
The author, who publishes a note daily using Claude Code subagents, breaks down the time per post: 19 posts were published from 9/22 to 10/10, at about 25 minutes each, of which /daily execution takes 10 minutes, image generation 5 minutes, and note pasting and publishing 10 minutes. In other words, 60% of the time is spent outside /daily, and image generation requires manually pasting the prompts in posts/ into nano banana pro, so it still cannot be fully unmanned.
The author used AI to generate a 4-cut short animation of a witch and a tiger cat's tea party in an autumn garden, which consumed 1,050 credits including revised versions (monthly limit 4,000), with cuts 2 and 3 each doubling to 350 due to revisions. The lesson is that the more elements there are, the more detailed the instructions need to be; for example, steam must be limited to "only from the center of the specified cup, thin and sparse, not blowing into the face," otherwise it steals the scene or appears in the wrong position. The author therefore changed to allocating seconds according to each scene's needs rather than fixing 5 seconds per cut, and only animating works with simple compositions, finally editing with CapCut and publishing to Instagram.
By drawing frames with Python, stitching with ffmpeg, and using the latest speech model only for narration, the cost of a 3-minute learning video can be reduced to about 50–90 yen, far below the 570–1,700 yen of video generation AI. The author recommends Gemini 3.8 TTS or Eleven v4 for narration, and says erroneous sentences can be re-read individually for easy correction.
Before generating AI renders, architects should first build a visible reference board in PureRef, assigning each image a single role—massing, material, light, or atmosphere—and grouping them while removing conflicts. PureRef is a desktop reference-image canvas that supports arranging, grouping, annotating, and exporting; it is not itself an architectural AI product. In ComfyUI, input only the needed reference images into IP-Adapter or image conditioning nodes, and record cropping and weights to avoid flattening the whole board into a single collage image.
To make a Product Hunt launch video, first identify the customer problem to demonstrate, then record the product solving that problem, and cut irrelevant details. Organize the video into a roughly 50-second story following "problem → product action → result → next step," and use a mood board to unify layout, pacing, and sound direction before motion design.
Veras 4.5's Remove Background can export any source image as a transparent PNG, but the official documentation publishes no benchmark tests for architectural cutouts or alpha precision. This feature tends to produce white, blue, or green fringes, broken lines, or hard chunks at the edges of architectural materials such as glass, railings, and vegetation; alpha edges need to be checked at 100% and 200% on pure white, mid-gray, and pure black backgrounds, and a decision made on whether to keep contact shadows. 4.5.2 fixed the issue of duplicate exports from sources with filters, rotation, or prior cutouts.
Kapwing published a step-by-step guide on using AI to generate Vox-style collage animation documentaries, broken into six steps: writing the script, breaking down visual beats, generating collage images, frame-by-frame animation, generating voiceover, and editing the final cut.
Why it matters: A complete breakdown of the six-step AI workflow for Vox-style collage documentaries, including reusable image and animation prompts, ideal for creators looking to replicate the visual style.
Creator sasao shared prompts for using the video generation AI "DomoAI" to create Zeppelin scenic videos, along with a complete English prompt example. The prompt describes a cinematic journey of a giant magical airship flying through clouds, skimming past floating islands and waterfalls, and finally soaring toward a floating kingdom in the sky, emphasizing no dialogue and no violence. The author also notes that DomoAI's free generation includes a watermark, recommends subscribing, and announces a new membership column with at least two updates per week.
AtomicChat released an AD-Q4_K quantized version of Qwen-Image-2.1-Turbo, sized at 4.2GB — roughly 70% smaller than the official BF16 version at 14.2GB. The author measured LPIPS to be 21% lower than the standard Q4_K.
Author Akita Ryosuke published his four-step workflow for using Claude desktop with Claude in Chrome to operate Seedance 2.5 for video generation: first, use ChatGPT to create a character sheet and prepare props and sound materials; then write out clear details for characters, scenes, actions, camera, dialogue pronunciation, negative prompts, aspect ratio, quality, and duration; next, let Claude automatically operate the Yapper platform to generate; and finally, provide feedback for revisions on each item.
Why it matters: The author published a complete four-step workflow for using Claude to operate Seedance 2.5 for video output, including character sheets, lip-sync details, and cost breakdowns — transferable to your own generation workflow.
An author used AI to adapt their note article about "selling things" into a 1-minute-28-second claymation, working from September 28 to October 9 and redoing it 9 times.
The author used MiniMax H3 in ComfyUI to fully locally generate a 2-minute-33-second Inkubus intro video. From conception to uploading to YouTube took about 6 hours, with GPU runtime of approximately 2 hours.
At tomi's request, the author used AI to turn their old story of riding a bike to buy oden ingredients in the autumn of their second year of middle school — when they couldn't brake on a downhill and had to hard-brake with their left foot — into a manga. The inspiration came from tomi's article about a 3m bicycle dive, along with related articles by ろくねこ and Kagaya. The author says AI manga is easier to draw than expected and that self-created characters are very convenient.