Trying Qwen-image3 img2img on a paid ComfyUI node
Qwen-image3 currently has no local version and must be called via API. The author topped up 5 USD on ComfyUI to obtain 1,055 credits for img2img testing, with each generation consuming 10 credits.
Qwen-image3 currently has no local version and must be called via API. The author topped up 5 USD on ComfyUI to obtain 1,055 credits for img2img testing, with each generation consuming 10 credits.
PROMPICKS, a Windows application for ComfyUI, has been updated to 1.2, adding prompt weight adjustment (0.1–10.0), restoring settings from generation history, one-by-one continuous generation and aborting, and raising the per-generation limit from 4 images to 32. The new version also supports saving tag combinations, pinning tags, and automatically saving selection states, and retains 5,000 standard tags, priced at 1,500 yen.
The author completed a 3-minute 26-second music video in a local environment with an RX 7800 XT (16 GB, ROCm) plus 64 GB of memory.
AIArtEngine released version v8.4.2, with a full changelog covering the changes from v8.4.1 to v8.4.2. The original text did not disclose specific features, parameters, or performance data.
The scenedetect project for ComfyUI has updated its recommended video node, which can now choose OpenCV or PyAV decoding and selects representative frames based on decoding timestamps, improving detection reliability for variable frame rate videos and short scenes. The PySceneDetect compatibility requirement has been raised to version 0.7.2 or higher, and PyAV can be optionally installed when not built into ComfyUI. A usage guide and example workflow for using VHS images with native video nodes have also been added.
Krea 3's Ultra mode was merged into the ComfyUI template on October 22, and the author tested the model's editing capabilities via the ComfyUI API. Krea 3 supports native 2K editing, region editing that redraws only changed areas, compositing of up to 8 reference images, HEX color value specification, and layout text preservation; Ultra has a fixed 2K output and takes several minutes per image.
The author optimized MOSS text-to-image in ComfyUI using the Spectrum node with an 8-step LoRA, testing 2-megapixel image output on an RTX3060 with 6GB VRAM and 16GB memory.
When using Z_image_turbo_bf16 to generate full-body portraits, the user found that facial close-up details were excellent, but the eyes appeared distorted and asymmetrical. After trying the two upscaling models RealESRGAN_x4 and 4x-ultrasharp, the results remained the same, and the positive prompt already included "Detailed eyes, sharp focus, detailed iris." The user asks whether there is a way to improve eye quality in full-body images.
The author used Animagine XL 4.0 Opt to separately generate the character and raindrop assets, then in ComfyUI used a mask of the transparent area of the window frame to composite the character and raindrops onto the inside of the glass, preserving the front-to-back relationship of the window frame being in front of the character.
Creator Mizuki shared their experience generating original character introduction posters with GPT, Grok, Meta AI, and Gemini; detailed step-by-step instructions repeatedly failed to meet expectations, and the final poster-style result was achieved through self-adjustment. Their daily generation environment is an RTX 3060 Laptop GPU (6GB VRAM) and 32GB memory, currently using a ComfyUI workflow built with ChatGPT and Grok, and they hope for a tool that can automatically generate high-quality video workflows suited to the local environment.
In the 2026 AI director software ranking, ScreenWeaver ranks first, followed by LTX Studio and Saga. The ranking evaluates six tools covering at least two production stages based on criteria such as "whether generated shots remain linked to the script text." ScreenWeaver integrates 7 video models including Seedance 2.5, Veo 3.1, and Kling 3.0; the Auteur version costs 39.99 USD/month, but the tool does not include full editing.
LTX production technical director Mohamed Oumoumad and Comfy's Doug Hogan explained in a livestream the professional VFX workflow LTX built for ComfyUI, aimed at real post-production pipelines and involving 16-bit EXR, ACES color, and delivery resolution output.
Pictory releases Pictory 3, adding an AI video agent: users describe their video requirements and answer five setup questions, after which the agent generates a script, scenes and a preview, and then generates the final video after revisions via conversation or the editor. Videos can start from an idea, script, webpage, document, PowerPoint or recording, with documents and URLs combined limited to at most 5 sources each time. New users will be moved into Pictory 3 with the rollout, while old projects remain in Projects and will not be converted automatically.
A user used Claude Code (Sonnet 5.5) with the Krea2_api workflow to generate more than 2,000 images in one night. Claude could freely call downloaded LoRAs, edit or create new workflows, and evaluate outputs and adjust parameters on its own, with the final notes.MD exceeding 1,700 lines. About 2,100 images along with the notes and summary files were packaged and shared in the comments.
A ComfyUI architectural rendering tutorial should first disclose the GPU and VRAM used for testing, the ComfyUI version, models and custom nodes, input dimensions, peak VRAM strategy and low-VRAM alternatives, and attach the workflow file and expected outputs.
The author used the MiniMax H3 Ref2VA model on local ComfyUI to produce a 10-minute comedy short drama, 7 types of vertical short videos and a 60-shot cat character video, at about 2 minutes per shot, or about 3.5 minutes after switching to a 768p 8-step LoRA.
Why it matters: The author ran MiniMax H3 on local ComfyUI to make multi-shot character videos, explaining one by one the causes of and fixes for broken hands, blinking and extra limbs, which can be transferred to your own production pipeline.
A user running the Z-Image WAN 2.2 image-to-video workflow downloaded from Civitai (WAN2.2_Img2video_auto.json) in ComfyUI kept getting "Bundle Auto-Loader failed".
Floo, a one-stop AI creation service platform, has officially launched, covering a script Agent, character design, a director's console, editing, and distribution revenue sharing. The platform integrates nearly 10 models including Opus5.5 and Seedance; the full version of Kling 4.0 debuts on Floo, at as low as 0.18 yuan per second. On the distribution side, it offers one-click agency for Hongguo Short Drama, Douyin, and WeChat Channels, and introduced a policy under which 100% of the revenue from works signed in Q4 goes to the authors.
Why it matters: Floo links script diagnosis, character setting, the director's console, and distribution revenue sharing into a single chain, presenting the complete process of AI short dramas from generation to monetization.
A user compared LTX 2.5 and Wan2.2 generating 5-second videos on an HP Omen 16 laptop (RTX 5090 with 24GB VRAM): LTX 2.5 took only 0.54 minutes, while Wan2.2 took 45 minutes. The user found that when running Wan2.2, the GPU used only 14GB of VRAM, with the rest offloaded to 32GB of system memory (28GB actually used); both setups used out-of-the-box default settings.
With WSL2's wslc capability, running ComfyUI on Windows no longer requires Docker or Python; simply run wsl --install in an administrator PowerShell and then run the wslc run command to start it.
When using a workflow with bounding boxes in the Moss image template in ComfyUI, the example knight keeps appearing in the prompt box and cannot be deleted directly. The user posted asking how to remove this example knight.
The MiniMax H3 Singularity V2 dual-sampling workflow has been updated to a new version, with a related demo video posted on YouTube. The workflow supports both Edit and Pruned modes; the post was reposted by user theOliviaRossi, who is not the author.
A user built a complete video generation pipeline in ComfyUI, from YuE2 generating songs, Z-Image Turbo producing character images, and Qwen-Image-Edit 2511 making keyframes, to Wan2.2-S2V lip-sync, two-stage rendering with Wan2.2-I2V, and FILM frame interpolation to 48fps. However, the generated results often show distortions such as three arms in hands and multi-person scenes, and the user is seeking suggestions for more realistic tools and open-source models.
ulearn launched the AI manga production platform Drawspire on October 15, which can generate manga storyboards from rough sketches and text descriptions, and detects whether images are excessively similar to existing works, with monthly fees in 4 tiers from 0 to 5,980 yen.
A user compared FastH3 Omni and Ref2VA + 8steps Turbo on a single 3090: all four shots were native 768×1344 at 24fps, with identical prompts, seeds, and identity and clothing references, and all backgrounds generated from text prompts. Omni averaged 8 minutes 26 seconds per render, while Turbo + VSA took 8 minutes 01 seconds, both covering 158-frame generation and showing the first six seconds of each segment. The user believes Turbo + VSA has better image quality and is slightly faster.
An AI anime studio, after using PixAI Studio and Tsubaki 3 to complete expression sheets, character emotional consistency, manga shading effects, and color-to-black-and-white manga storyboard conversion, is now looking for reusable ComfyUI workflows. Its self-built solution relies on character LoRA, IPAdapter, and ControlNet, but maintaining character consistency across outputs is very costly, and in production it prefers ready-made workflows.
The V2 INT8 checkpoint of the community fine-tuned model MiniMax H3 Singularity was uploaded on October 21 to the WarmBloodAban/Minimax-h3_Singularity repository, with the file Minimax-h3_Singularity_ref2va_v2_int8.safetensors, size 31.7 GB.
Stable Diffusion Art released a tutorial on how to use MiniMax H3's Add Guide node in ComfyUI to lock multiple keyframe images, video clips, and audio to specific frame positions on a 24fps timeline, generating a video with stereo dialogue, sound effects, and music in one pass.
Why it matters: The tutorial provides specific node settings for multi-keyframe locking and audio-visual continuation, and fixes the timestamp truncation issue in the official template.
Creator kingroka released the LoRA "UDA" for use with Moss Image to generate anime and digital paintings, to alleviate the strange artifacts and errors that the base model produces in digital art. The LoRA is based on about 250 images, trained for 10,000 steps, with the dataset using structured JSON bbox prompt types, and incorporates merged conversion data from three Ideogram LoRAs: anime/digital painting, Ghibli style, and retro anime.
DramaClip v0.2.2 is released, adding three toolbox capabilities: video copy extraction, speech-to-text, and text-to-speech, and includes a built-in HTTP gateway with local loopback only plus token authentication for external Agent calls.
Xiaozhi Blog released a tutorial introducing how to do text-to-video with MiniMax H3 in ComfyUI, covering environment preparation, hardware and software requirements, and obtaining MiniMax H3 model files from ModelScope or Hugging Face, and provides steps for installing and upgrading huggingface-related dependencies.
Longshot Studio has been updated to 0.15.0, with multiple rounds of iteration on lip-sync and facial refinement; the API is free and runs on the user's own ComfyUI, and the code is open-sourced on GitHub. The author also released a music video, "Falling Fast," produced entirely with the studio as a demonstration.
A user built a comparison page testing the image editing capabilities of Fibo 2 edit, Klein 9b, and Qwen 2.1 at 1 megapixel with the same seed and int8 single-pass generation, concluding that Qwen 2.1 remains the best for their use case, and Fibo could be a good choice if it gains community support.
HeyGen Voice Turbo integrates into ComfyUI as a partner node, supporting voice cloning from a single recording and converting any script into speech. Turbo costs less than 1 cent per 1,000 characters, targeting scaled voiceover scenarios such as localized ads, training videos, product explanations, and AI digital human pipelines.
Open-source canvas Agent released v1.6.9, unifying the schema, Policy, visibility, execution binding, and approval preview of 29 tools, and removing old tool copies and duplicate context pipelines. Canvas Agent fixes focus loss when more than 8 nodes are multi-selected, and step cards add step-by-step undo; submitted generation tasks cannot be undone. The backend database Schema is upgraded to 58, adding Agent memory compression proposals and revision records; upgrading requires updating both Web and Backend.
User Sixhaunt trained a rank-16 LoRA (6,000 steps) with AI toolkit and compared two reference-image input methods for Moss Image Edit: the direct reference-image method tends to mix inputs together without a LoRA, while Ostris's editing method approximates image overlay without a LoRA and gives better editing results with a LoRA. The Ostris method renders about 45s per image, the simple reference method about 15s, and training takes about 3 times as long.
MattVidPro used Claude Opus 5.5 to rebuild fictional game footage generated by Seedance into a playable browser shader in 45 minutes, with the help of gokaygokay's video-to-shader skill.
MiniMax H3 Novel Pipeline released an update, adding editable reference links and a review process, and supporting recovery of extraction from saved cinematic chapter JSON. The H3 generation stage adds scene duration guidance, configurable maximum shot count, and optional shot direction, and can use Qwen to correct invalid shot post-processing notes based on validation feedback.
The author demonstrates completing a real generation in local ComfyUI without relying on a local inference GPU and without downloading models, by installing the free Forma custom node and connecting an account, then opening the image-to-video template. The solution is pay-as-you-go with no subscription, top-ups starting at $5, with 39 image and video models, 8 video workflow types, and 2 image workflow types to choose from, and quotes can be viewed before running inference.
The author shared a 360-degree orbit MiniMax workflow in ComfyUI, calling it the definitive method they found for multi-angle face consistency. The approach is to first prepare a very solid frontal face image, then have the AI do a 360-degree orbit and capture frames at each important angle; the author claims face consistency reaches 100%, outperforming the search for real photos of real people at different angles.