ComfyUI’s cover photo
ComfyUI

ComfyUI

Software Development

The most powerful open source workflow engine for visual AI

About us

Comfy is the world's leading workflow engine for visual AI. Powered by ComfyUI, the open-source platform trusted by 4 million creators across 118,000+ GitHub stars and 60,000+ community-built nodes, Comfy gives creative and technical teams complete control over every model, every parameter, and every output. From independent artists to enterprise studios, professionals at Amazon Studios, Apple, Netflix, Nike, Ubisoft, and Pixomondo trust Comfy to power their AI-driven pipelines across VFX, animation, advertising, gaming, and eCommerce. Comfy's node-based canvas makes every AI decision visible, inspectable, and reproducible. No black boxes, no locked ecosystems. Whether you run workflows on your own hardware with Comfy Desktop, deploy at scale with Comfy Cloud, or integrate directly with Comfy API, you own the output. Backed by $47 million in total funding at a $500 million valuation, with investors including Craft Ventures, Pace Capital, and Chemistry.

Industry
Software Development
Company size
11-50 employees
Headquarters
San Francisco
Type
Privately Held
Founded
2024

Locations

Employees at ComfyUI

Updates

  • ComfyUI reposted this

    The prompt box was never the product. It was the demo. Three years of generative AI tooling optimized for the first ten minutes: type a sentence, get something impressive, feel the future. Then you need it twice. Then you need a specific model version, a control signal, a step inserted between two other steps, an output that matches the one from last Tuesday. The box has no answer for any of that. Everything past the demo becomes a support ticket. A node graph has an answer, because a graph is just the pipeline with the lid off. Here's the part I think the market keeps missing. When you build on the pipeline instead of hiding it, you stop being a product for one modality. The same canvas that runs text-to-image runs video, audio, 3D, upscaling, batch pipelines, and whatever architecture drops next month because a graph doesn't care what's flowing through it. That isn't a feature we shipped. It's what falls out of the architecture. Every closed generation product has to rebuild itself each time the frontier moves. We add a node. And I don't accept that visible complexity and approachability are opposites. That's a design problem, not a law of physics. Reroutes and Get/Set nodes kill the spaghetti. Subgraphs collapse a hundred nodes into one. App Mode wraps the whole thing in a clean front end for the people on your team who should never see the wiring. Same backend, three levels of exposure. You pick which one you're standing at today, and you can move between them without changing tools. That's the bet: depth you can grow into rather than a ceiling you hit in week two. Module 1 of the ComfyUI beginner series is out - full UI tour, the core nodes underneath every workflow, and how to stop being intimidated by the canvas. Runs on Comfy Cloud, nothing to install. The market is splitting into things that generate media and infrastructure people build companies on.

  • Seedance 2.5 vs. MiniMax H3 with 3D blockouts. Across these tests, Seedance tends to pick up more natural movement and follow the blockout a bit more closely. But it’s not a clear difference every time, some generations are nearly 1:1 between the two models. MiniMax becomes especially interesting when cost and iteration are part of the equation. In these 5-second Comfy Cloud tests, it ran at a fraction of the credit cost of Seedance, and with the right hardware, it can also be run locally for free. Both models have their place. The better fit comes down to the use case, the level of motion and control you need, how much you want to experiment and iterate, and the budget you’re working with. Sometimes the extra capability from Seedance is worth the additional cost. In other cases, getting close with MiniMax while having significantly more room to iterate can make more sense. Want to try this workflow yourself? Link in the comments below. MiniMax H3: https://lnkd.in/gDRVRk6D Seedance 2.5: https://lnkd.in/gmugMb-e

  • ComfyUI reposted this

    The generative AI argument in film and television is over. Amazon Studios posted a Gen AI / VFX Coordinator role last week, and it names seven tools. Six are closed APIs. ComfyUI is the one the artist runs on their own machine. Which means the studio is committing to a workflow it doesn't rent. That's a different kind of decision, and it's the one I'd pay attention to. The rest reads like infrastructure. Culver City, reporting up to the Director of Visual Effects for a US Series. A per-show checklist for security approvals, tools approvals, and Legal Gen AI use-case approvals. An internal database of global AI vendors and their capabilities. That's not a studio testing whether this works. That's a studio building the compliance layer you only build once you've decided it's permanent. There's a line item for gathering availability of Freelance AI Supervisors. A craft role that didn't have a name three years ago. And the scope covers final pixel, not just previz. Concept art and pre-production were always the safe answer, because nothing there ships. Final pixel goes in the episode.

    • No alternative text description for this image
  • Video Model Showdown: Open-Source vs. Paid AI Video Models It's a Video Model Showdown: Purz and Allyson put two of the newest AI video generation models head-to-head, running the exact same prompts side by side to see how they actually stack up. Does price really = quality? We'll generate the same shots on both models live, compare motion, fidelity, and prompt adherence in real time, and dig into the question everyone in the ComfyUI community is asking: does paying more for a closed model actually buy you better output, or is the open-source option already closing the gap? Bring your own prompt ideas and we may run them live on both models!

    Video Model Showdown: Open-Source vs. Paid AI Video Models

    Video Model Showdown: Open-Source vs. Paid AI Video Models

    www.linkedin.com

  • ComfyUI reposted this

    View organization page for LTX.io

    14,787 followers

    ComfyUI builds for people who want to own their pipeline end to end. No waiting on a roadmap. No renting access by the call. Just a graph you wire yourself and control you keep. That's the belief driving how ComfyUI works with LTX. Our models stay open, so Comfy's graph updates as they evolve, and builders never lose control of their own workflow.

  • A music video test in MiniMax H3, built entirely from generated stills. Five reference images, one for the subject, four for the locations, went in alongside an audio track as a single context. H3 is an omni-modal model: text, images, video, and audio go into one context rather than through separate conditioning paths. That is what makes this kind of shot list possible. Identity holds across every reference, environment and wardrobe come from the location stills, timing comes from the audio. The clip comes back with the location cuts already in it. Lip sync is the inconsistent part. Run to run it either syncs or it does not, and that is the thing to budget takes for. Open weights landed August 3, with native ComfyUI support merged the same day. Open weights also mean the community can quantize it, fine-tune it, and build nodes around it rather than waiting on an endpoint. Want to try this workflow yourself? Link in the comments below. https://lnkd.in/gFhtTP49

  • View organization page for ComfyUI

    43,900 followers

    LTX-2.5 is native in ComfyUI, day zero with the weights LTX-2.5 isn't a version bump. The LTX.io team rebuilt the generation stack rendering approach. New decoder. New text encoder. New base checkpoint. LTX-2.5 allocates compute by complexity instead. Motion, composition, and framing generate first in a compressed latent space alongside a set of high-fidelity keyframes more for complex scenes, fewer for simple ones. A dedicated pixel-diffusion stage then renders the final video from both. What that changes in practice: → Textures, materials, and faces resolve with pixel-level precision, and busy shots automatically draw more rendering compute than static ones → A Diffusion Video Decoder replaces standard VAE decoding - sharper faces, legible text, fewer smears in fast motion → Native multi-shot holds character, environment, lighting, and voice across cuts, so a sequence comes out of one run instead of being matched together afterward → A custom Gemma 4 12B text encoder keeps long prompts intact instead of dropping clauses Three variants ship: the main model, a reworked distilled version, and a non-SFT pretrained checkpoint for aggressive fine-tuning. Native 4K, synced audio and video, and up to 50fps carry over from 2.3. Learn more and download the workflows with the link below 👇 https://lnkd.in/gpn76_qB

  • Comfy Creatives Model Jam: MiniMax H3, Seedance 2.5, Wan Animate 2 & More It's a Model Jam! The Comfy Creatives team is hopping on live to mess around with a stack of brand-new model releases. No script, no pressure, just seeing what these things can actually do. On deck: 🎨 MiniMax H3 🎨 Grok Imagine Image 2.0 🎨 Qwen Image 3.0 🎨 Wan Animate 2 🎨 Seedance 2.5 🎨 Seedream 5.0's new layer separation feature We'll be generating live, comparing notes in real time, and following whatever's fun or weird. Come hang out, watch the models in action, and throw your prompts and questions in chat! ⏰ Subscribe and enable notifications to catch the stream live 💬 Submit your questions in chat during the broadcast or before the stream in the comments! 👍 Leave a like if you'd like to see more content on AI + post-production workflows!

    Comfy Creatives Model Jam: MiniMax H3, Seedance 2.5, Wan Animate 2 & More

    Comfy Creatives Model Jam: MiniMax H3, Seedance 2.5, Wan Animate 2 & More

    www.linkedin.com

Similar pages

Browse jobs