October 2026 Summaries
9 posts from Comfy
Filter
Month:
Year:
Post Summaries
Back to Blog
ComfyUI has announced a two-week competition offering $25,000 in prizes for open-source creative applications built on its Developer Platform, framed as an effort to encourage publicly available alternatives to tools such as Genjutsu and Cinema Studio. Participants must publish repositories that others can run, modify, and improve, with awards including a $10,000 grand prize, RTX 5090 GPUs for the most innovative, most viral, and open-weights-only projects, and Comfy Credits for 10 runners-up. Additional partner bonuses are available for workflows centered on LTX video, FLUX3, or MiniMax H3 models. The competition is backed by NVIDIA AI, Runpod, LTX, Black Forest Labs, and MiniMax, and ComfyUI is also hosting in-person build events in San Francisco and Los Angeles on October 28 and 29.
Oct 15, 2026
393 words in the original blog post.
FLUX 3 Image Klein is a 5.6-billion-parameter open-weights image-generation model from Black Forest Labs that runs natively in ComfyUI and is offered in base and distilled versions for research, fine-tuning, or faster generation. It retains advanced controls from the larger FLUX 3 Image model, including native 4K generation, prompt expansion, multi-reference editing, bounding-box-based composition control, text rendering and editing, and reference-driven style transfer. Black Forest Labs reports that it outperforms Qwen-Image 2.1 in internal text-to-image and multi-reference evaluations, while INT8 ConvRot quantization combined with the distilled model can make 1024×1024 generation 4.3 to 9.3 times faster than a BF16 50-step baseline, with reported image times falling below nine seconds on RTX 5070 and RTX 4070 GPUs. Users can access it by updating ComfyUI or using Comfy Cloud, downloading the preferred weights from Hugging Face, loading the supplied template, and providing prompts or reference images.
Oct 15, 2026
366 words in the original blog post.
Krea 3, Krea’s reference-driven image generation and editing model, is now available in ComfyUI through Partner Nodes, enabling users to work from starting images and multiple visual references while specifying both desired changes and details that should remain unchanged. The model emphasizes rapid, lower-cost iteration, supports numbered references for assigning specific roles such as a person, garment, or background, and can make targeted material, color, lighting, composition, and surface edits while preserving features like labels, object shapes, camera angles, and materials. Users can reuse successful outputs as new inputs to refine imagery or create sequences for later video workflows. To begin, users update ComfyUI or use Comfy Cloud, locate the Krea 3 node or template, sign in to a Comfy account, and run the Partner Node using Comfy credits, with recommendations to clearly identify preserved elements, define each reference’s purpose, and return to stronger earlier results if a sequence begins to drift.
Oct 15, 2026
447 words in the original blog post.
Nvidia’s RTX video effects are now available as custom ComfyUI nodes, enabling VFX professionals with RTX GPUs to locally upscale footage, generate interpolated frames, and convert SDR material to BT.2020 HDR. The recommended workflow chains Video Super Resolution, which offers up to 4× scaling, denoising, deblurring, and multiple source-specific quality modes, with Frame Generation for 2× to 16× frame-rate expansion or custom-timed interpolation, followed by HDR conversion with adjustable contrast, saturation, middle gray, peak luminance, and debanding. The release also improves ComfyUI’s VFX-oriented EXR workflow through support for half- and full-float EXRs, preservation of superwhite and negative pixel values, expanded color-space conversion options, JPEG previews and HDR Viewer access for EXR assets, and simplified downloading of image sequences. Users can update ComfyUI, install the Nvidia VFX nodes through ComfyUI Manager, and access included example workflows for upscaling, interpolation, HDR conversion, and combined VFX processing.
Oct 13, 2026
1,032 words in the original blog post.
ComfyStreamerH3 is an open-source experiment aimed at producing roughly 15 seconds of watchable 448×256 AI video in real time or faster on a single RTX 5090 GPU, addressing the high cost of continuously running larger video-generation systems. Built around FastVideo’s FastH3 V2 checkpoint, the workflow combines four-step sampling, sparse attention, a smaller Qwen3-VL-based text encoder, a pruned INT8 model checkpoint, fused FP4 MLP operations, and overlapping tiled decoding with NVENC encoding to reduce memory needs from about 80 GB to under 30 GB. While resulting clips exhibit artifacts and limited detail, the project demonstrates that partially legible live video can be generated on consumer-accessible hardware rentable for about 70 cents per hour. The author released the custom ComfyUI node through GitHub and suggests that inexpensive post-generation upscaling, such as SolRefinery, could improve resolution and make single-GPU live video more practical in the future.
Oct 08, 2026
884 words in the original blog post.
Google’s Gemini Nano Banana 2.1 image generation and editing model is now available in ComfyUI through Partner Nodes, offering lower-cost output at resolutions from 1K to 4K and compatibility with existing Nano Banana 2 workflows. The model supports Minimal, Medium, and High thinking levels that trade speed and token use for greater ability to handle dense prompts and complex layouts, along with up to 14 reference images for combining products, characters, styles, and backgrounds. Its features include Google Search grounding for real-world details, improved rendering of specified text and multilingual content, multi-turn targeted edits, seamless tile generation, and a wide range of aspect ratios including panoramic and vertical formats. Users can access it by updating ComfyUI or using Comfy Cloud, then selecting the node or a template workflow, configuring output settings, adding prompts and references, and running the generation.
Oct 06, 2026
380 words in the original blog post.
AI animation artist 852話 Hakoniwa collaborated with the ComfyUI team to create YUI, described as the first animated short film made with Comfy Agent, combining a narrative about a girl facing the end of her world with a behind-the-scenes account of its production. Hakoniwa, formerly an illustrator, Vocaloid creator, and indie game developer, adopted AI tools after discovering Midjourney in 2022 and has since become an early user of open-source models and ComfyUI workflows while co-running Tokyo-based AI company AndIdea. YUI, whose name evokes both ComfyUI and the Japanese concept of connection, evolved from an earlier red-haired character named HIBANA into a yellow-and-purple protagonist designed with GPT Image 2.5 and guided by character sheets, storyboards, and hand-drawn compositional sketches. The film follows YUI as a timer announces that her world will end in three minutes, though it freezes at one second and extends the story into the production process itself. Hakoniwa completed the project over roughly 7 to 10 days, including three focused days of individual work and approximately 200,000 yen in costs excluding labor, using Comfy Agent to regenerate shots, compare video models and upscalers, check character consistency, and reduce repetitive workflow tasks while retaining creative control over pacing, expressions, composition, and final selections.
Oct 01, 2026
1,181 words in the original blog post.
ComfyUI is hosting the Comfy Dev Platform Challenge from October 5 to 19, inviting developers to recreate paid creative AI features such as cinematic controls, character consistency, relighting, and face swapping as open-source tools. Participants must build with the Comfy Developer Platform, Comfy Cloud, or Comfy SDK; publish a licensed, reproducible repository; provide a demo video; and comply with rules requiring lawful, safe-for-work work without unlicensed intellectual property or likenesses. Entries can use Comfy API, SDK, and Router to deploy workflows and access image, video, 3D, and audio models, with prizes including a $10,000 grand prize, NVIDIA GeForce RTX 5090 GPUs for practical, entertaining, and fully open-stack workflows, plus model-specific API-credit bonuses for LTX and Flux projects. Judges from ComfyUI, NVIDIA, and Runpod will assess submissions for practical usefulness, completeness, reproducibility, delight, shareability, and, for entertainment entries, demo quality. The challenge also includes free credits for the first 100 registrants, community support through Reddit and Discord, and an October 13 San Francisco event for builders and model labs.
Oct 01, 2026
955 words in the original blog post.
ComfyUI has introduced Comfy Agent, a beta AI assistant available in Comfy Cloud and planned for Comfy Desktop, designed to help creators turn natural-language ideas into working image and video generation workflows. The agent can plan, build, run, explain, revise, and troubleshoot workflows while drawing on current models, extensions, assets, and techniques, allowing users to focus more on creative direction than node wiring or technical terminology. It works collaboratively on the same canvas as the user, supports visual asset understanding, public or private skills, and up to five parallel chats with saved histories, while retaining users’ access to every workflow parameter. Early demonstrations include character animation, scene exploration, product variations, and VFX tasks such as masking, depth, normals, and compositing passes. Users can try it through the Agent panel in Comfy Cloud using existing Comfy Credits, while future plans include local GPU and model support, hardware-aware setup assistance, agent-built custom nodes, expanded model choices, integrations, and improved skills.
Oct 01, 2026
697 words in the original blog post.