A user used an inpainting workflow to compare MOSS-Image, Flux.2 klein 9B and Qwen Image Edit (2511) side by side, finding that MOSS-Image underperformed expectations, while Flux.2 Klein and Qwen Image Edit could reliably complete the task. The user added that MOSS-Image performs well in regular text-to-image (T2I), and acknowledged that the test may involve skill factors; the workflow will be posted in the comments.
ComfyUI Comic Creator has been updated to v1.62.0, adding an OCR tab and a re-read feature based on PaddleOCR-VL-manga; currently the backend only supports Ollama. Users can install the model via `ollama pull hf.co/LunarOilRig/PaddleOCR-VL-manga-GGUF:Q4_K_M` and generate images in the moss-image tab.
Higgsfield released a set of Seedance 2.5 native 4K tests on October 24, including using a Blender blockout to directly drive the generation of a bee scene, and comparing skin texture between direct 4K output and upscaled output.
A Reddit user compared the outputs of 7 image models using the same portrait prompt, ranked in order: Nano Banana 2.1, GPT Image 2.5, Muse Image, Qwen image 2.1 turbo, Flux 3 Klein, Krea 2 turbo and Moss-image-sft (8 steps LoRA).
ArcReel v0.34.0 has been released. For agent providers without pricing tiers, models can be selected manually; credentials support setting multipliers, and subscription-based gateways set to 0 are not billed. Billing now settles according to the model catalog price list, with text, voiceover, image and video costs recorded separately, and unpriced calls counted separately. The model catalog adds a search interface and syncs prices and capabilities with each release; Seedance 2.0 offers 4K and Seedance 2.5 offers 1080p.
AI Video Production Editor v3.1.0 has released a desktop installer, adding one-click connection for Claude & Codex. The studio runs a local MCP server and a REST API containing 68 tools; agents can operate stages such as script, characters, director, storyboard and generation, with changes displayed in the studio in real time.
For the 2026 need from script to picture, ScreenWeaver, Saga, Morphic, Autodesk Flow Studio, Higgsfield Popcorn and Google Flow are listed as six alternatives to LTX Studio.
A user asks for help on how to edit videos with H3 Minimax or LTX to adjust lighting or add effects. Their usual process is to use the original video as the animation source and image 1 as the visual end state, handed to Minimax for editing; it works for medium and close shots, but for wider shots, even at 2 mp, 30 steps and without turbo, facial animation is completely lost, and trying various sigmas and samples as well as face refiner was ineffective.
A ComfyUI user shared their image prompt enhancement workflow based on LM Studio, testing 7 GGUF models from 4B to 35B and finding that some models performed poorly. One GPT model, even when the system prompt and Structured Output explicitly required no reasoning, still output reasoning content, while other models were normal. The user said they are unwilling to pay for cloud services and mentioned the results of gemma 4.
Under the same prompt, Claude Opus 5.5, Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017 were compared on generating a "riding and flying" scene; Kimi K3.1 Pro integrated most stably, with the bicycle steadily riding on the flying platform, clear characters, clean lighting, and noticeably higher scene completeness.
The author tested generating a 30-minute film using the Minimax H3 open-source model with a self-hosted Atlas Cloud endpoint (2.4 cents/second), bringing the cost down to under 100 USD, whereas a previously produced 31-minute vertical film made with Seedance 2.5 cost about 600 USD in total.
ArchiGen proposes using a "preparation ledger" to compare AI renderers: timing starts when the source project is opened, not when the AI task starts, and must record manual minutes for model cleanup, camera setup, source image rendering, control channel export, mask and reference image selection, etc. Reusable setup and per-view work should be counted separately; for example, 45 minutes of one-time setup plus 5 minutes per view beats 2 minutes of setup plus 15 minutes per view after the fourth view. Failed exports also count toward work hours, and finding reference images and checking licensing are also input labor.
Chaos released Veras 5.3 NODES on October 19, turning the node graph into a managed product: the node catalog, cloud execution, and Revit, Rhino, and SketchUp output are all handled by the vendor, and settings can be pinned to the CONTROLS view, hiding the canvas from the operator.
A review of MiniMax H3 Light Cone Edition VBVR shows enhancements in three areas: prompt adherence, spatial relationships, and physical logic. The review centers on the version's actual generation performance, focusing on verifying the improvements in the above capabilities.
NVIDIA upgraded RTX Remix to version 1.6 (GitHub shows 1.6.1), adding experimental path tracing support for Minecraft Java Edition 1.21.1 and introducing DLSS 5's 3D-guided neural rendering and up to 6x multi-frame generation.
A user who migrated from WebUI to ComfyUI found that the KSampler node runs all steps regardless of the denoise setting, and only manually setting the steps to "original steps × denoise" produces the expected effect. They asked whether this is a change in node behavior, and whether this multiplication is correct or requires a more complex conversion, and are considering using a simple node to automatically calculate steps based on denoise.
A ComfyUI user generated NSFW images with Anima plus Anisee V2 Asthetic and posted asking whether there is a platform where works can be uploaded with the generation method made public, so workflows can be reused and image quality improved. Someone in the thread recommended Civitai red, and the author later confirmed and thanked them.
The author used Hermes Agent with Qwen3.8 Flash Next 125B IQ3_S and MOSS-Image on ComfyUI to generate Japanese infographics and manga-style images locally.
The author trained a LoRA for MOSS-Image; combined with the RGBA VAE from trmz/flux2-klein-alpha, adding isolated on a transparent background at the end of the prompt yields a transparent PNG directly from the sampler, with no need for matting or chroma keying.
A photographer built a custom ComfyUI node, "load image recursive," so the built-in Load Image node can recursively list images at any depth under input/, and added a folder dropdown filter, shrinking a scrolling list of 1200 entries to 40.
The author built a personal character design Skill with GPTimage, fixing the art style to GPTimage's mainstream style and lowering the head-to-body ratio so characters lean more toward animation proportions. In testing, three comparison sets of a dog-eared maid were generated using only prompts, a conventional art style, and head-to-body adjustments, and the memoryless GPT version still produced dog ears that were not accepted. The author says the Skill is reusable and can also make adjustments without regenerating.
VideoGen released its first in-house video generation model, Model 1, on October 20; it can generate multi-shot videos up to 10 minutes long with character consistency across scenes from a single instruction, referencing up to 16 images, and is currently available to select partners.
A creator compiled Midjourney prompts into collections and sold them on note, Gumroad, and PromptBase, achieving a stable monthly income of over 50,000 yen. On note, the "Anime Character Generation Prompt Collection" is priced at 1,000 yen and sells an average of 150 sets per month; the English version on Gumroad sells about 100 copies per month at 15 USD each; and PromptBase sells about 150 copies per month at 5 USD each.
A user trained multiple character LoRAs for Moss Image using ai-toolkit and DiffSynth, but when generating images, every person in the image shows a variant of the same face from that LoRA. They tried bounding boxes and various prompting methods with no effect, and as a LoRA beginner are asking what the problem is.
The red tribute micro short drama "I Have Seen Your Journey" launched on October 22 on platforms including Hongguo Short Drama, Douyin, and Z View, and will later air on Zhejiang Satellite TV; it is the first Zhejiang Satellite TV "Taihua Mark" micro short drama project. The drama is among the first batch of selected works for the National Radio and Television Administration's "Long March: Our Story" themed micro short dramas, focusing on the historical facts of the 1934 lone expedition of the Northward Anti-Japanese Advance Team and its coordination with the strategic transfer of the Central Red Army; it is co-produced by Rongliang Short Drama and others.
Training of ByteDance's Goku video model crashed frequently in mid-2024; engineers eventually traced it to tampered instructions in the deep learning framework that reordered inputs by time, making failures impossible to reproduce, and access logs pointed to Peking University intern Tian Keyu. The company found a hidden directory in his account and programs suspected of matching the data pipeline sabotage, and launched a formal investigation in early August.
This prompt collection provides approximately 100 JSON-format image generation prompts that can be directly copied into tools like ChatGPT and Gemini. Centered on everyday scenes such as coastlines, old bookstores, train stations, airports, and street corners, the prompts use perspective, reflections, transparent materials, lighting, and a sense of scale to create visuals that "turn the everyday into the extraordinary."
The National Radio and Television Administration issued the "15th Five-Year Plan for Radio, Television, and Online Audiovisual," proposing that by 2030, radio and television media will fully advance into the main battlefield of the internet. China Media Group launched the "Multilingual Short Drama and Anime AI Production and Publishing Platform," leveraging AI large models and multilingual corpus resources to explore AI-empowered audiovisual creation, and simultaneously released the AI anime project "Lia: Meeting the Stars." Kuaishou Kling AI fully released the Kling 4.0 video generation model, supporting 4K and 1080p 10-bit HDR output, with a single generation up to 30 seconds.
Bona Film Group's 100-minute AI hyper-realistic feature film "Sanxingdui: Future Past" has obtained a public screening license from the China Film Administration, becoming the first domestic AI feature film to reach theaters. The film was completed by a team of 70–80 people over more than two years, with a total investment of over ten million yuan, and the team generated purely digital human characters directly without any live-action shooting throughout. A previously released AI short drama on the same subject, "Sanxingdui: Future Revelation," received over 200 million views on Douyin, providing a test for the feature film direction; Bona's second AI film, "Off-Plan: Guanghan Operation," has already started production.
Why it matters: Bona spent two years and tens of millions of yuan to run through the production and review process of its first AI theatrical feature film, providing a reference for small and medium teams to evaluate the cost structure of AI feature films.
The author used Animagine XL 4.0 Opt in ComfyUI to generate a "reverse sitting on a chair" character, relying solely on tags to specify the reverse sitting pose, backrest in front, arms resting on the backrest, and forward-leaning posture, without using LoRA, ControlNet, or input images.
How is story-driven AI shaping the future of filmmaking and the entertainment industry? At this year's Busan Asia Content & Film Market (ACFM), Kling AI joined top filmmakers and industry leaders from across Asia to delve into this question. Come take a look!
A ComfyUI user searched in Documents and ComfyUI-Shared\models but still couldn't find the custom_nodes folder, preventing them from installing ComfyUI Manager. The user said they were following a tutorial and needed the folder to continue.
The author, not wanting to use ComfyUI's node workflow, used ChatGPT to build a browser frontend called MOSS Image Studio, with the backend still running ComfyUI. The frontend supports MOSS-Image text-to-image, experimental image-to-image, MOSS Base/SFT and INT8/BF16 variant selection, as well as resolution, steps, sampler, and other parameter settings, plus generation queue, project saving, and workflow export.
At the DataEye L!NK 2026 Short Drama Going Overseas Global Growth Conference, industry experts from Maiya, Wondershare Technology, Llinks Law Offices, Chuangxiang Shuwei, and Airwallex (空中云汇) dissected the full chain of short-drama going overseas. Guests pointed out that strong emotional hooks can compensate for insufficient localization, and going overseas requires two-way matching of culture and content, as well as building copyright and data compliance defenses. Data assetization and payment risk control (subscription checkout system, chargeback risk management, local payment capability) are key to securing revenue.
The author locally produced a 4-minute unofficial fan MV for a The Flaming Lips song using a single RTX 4090 and open-weight models, without any cloud services or paid APIs throughout.
Why it matters: The author published the complete pipeline and per-stage timing for completing a 4-minute MV on a single card locally, which can be used to estimate the computing power costs of building a music video workflow.
A user who returned to AI video creation from the Disco Diffusion and Stable Diffusion era said they can't stand tutorials hosted by AI-generated voices and AI-generated characters, and would rather see real people occasionally stumbling or going off-topic. They also grew tired of the exaggerated-expression AI tutorial thumbnails on YouTube, and said they posted this rant while waiting for rendering.
VideoGen launched the Model 1 Creator Program, where any film using Model 1 in production can participate. By crediting "Made with VideoGen Model 1" and submitting, participants can receive million-level email pushes, social media exposure, extra credits, and media resource connections.