Best ComfyUI image models (2026): from SD 1.5 to Qwen-Image 2.1

22 Oct 2026 | AI – Artificial Intelligence, AI Tutorials, ComfyUI

This is part of our AI and ComfyUI tutorials. For every model you'll find a link to its official page, where you should always check the current licence. Our guide to models for creating video has its own page.

Every time we open Hugging Face there's a new image model, and every time we ask ourselves the same questions: is it worth downloading? Will it run on our 12 GB RTX 3060? And above all: can we use it for paid work? That's where this guide comes from. It's the cheat sheet we use ourselves.

Stable Diffusion, SDXL, Flux, Qwen-Image, Z-Image, LongCat, ERNIE-Image, Ideogram, Boogu, Krea 2… Here we put them in order, from oldest to newest: who made them, what they're good at, how much video memory they need and what their licence allows. Because "free to download" doesn't always mean "free to use for work".

All the models in one table

This page is a plain-language guide, not legal advice. Licences can change: before using a model for a job, always read the full text on the official page linked in each section.

Licences explained in plain words

Every model on this page is free to download, but the licence decides what you can do with it. We always separate two things: the model (the file you download and run) and the images you create with it.

• Apache 2.0 and MIT — the most permissive: you can use both the model and the images for work, modify the model and build it into a product, as long as you keep the copyright notices.
• OpenRAIL (Stable Diffusion 1.5 and SDXL) — commercial use is allowed, but there's a list of prohibited uses (illegal content, defamation, disinformation and so on) that you must respect, and pass on if you redistribute the model.
• "Community" licences with a threshold (SD 3.5, Krea 2) — free for work too, as long as your business stays below an annual revenue threshold ($1 million for both); above that you need a paid licence.
• Non-commercial, but images are free to use (FLUX.1 Dev, FLUX.2 Dev, FLUX.2 Klein 9B, Anima) — the model itself is for non-commercial use only, but the licence states that the generated images can be used commercially. Without a licence from the developer, you can't use the model itself in a paid product or service.
• Non-commercial only (Qwen-Image 2.1, Ideogram 4.0) — no work use without a separate licence from the developer. Be careful with Ideogram 4.0: the images stay yours, but the free licence specifically excludes generating images for revenue-generating products.
• Geographic exclusions (HunyuanImage 2.1 and 3.0) — Tencent's licence explicitly says it does not apply in the European Union, the United Kingdom or South Korea. If you live in the EU, there is simply no licence covering your use of the model.

How much VRAM you need

Your graphics card's memory (VRAM) decides which models you can use. In ComfyUI almost every recent model also comes in reduced versions: fp8 (about half the memory, nearly identical quality) and GGUF (even lighter, with a small loss in quality). The figures in each section are rough guidelines for ComfyUI with these versions.

• 6–8 GB: Stable Diffusion 1.5, SDXL and Anima; some recent models only in GGUF, and with patience.
• 12 GB: FLUX.1 in fp8 or GGUF, Z-Image Turbo, FLUX.2 Klein 4B, Chroma, Ideogram 4.0 in a reduced version.
• 16 GB: almost all mid-sized models in fp8, with room to spare.
• 24 GB and up: the biggest models (Qwen-Image, HiDream, FLUX.2 Dev) with fewer compromises.

If you need to pick a graphics card, read our guide to the best GPUs for AI. If you don't have ComfyUI yet, start with our guide on how to install ComfyUI on Windows.

The first generation: Stable Diffusion

Stable Diffusion 1.5 (2022)

The model that made it possible to generate images on your own computer. Released in October 2022 by Runway together with Stability AI, it's small (about 1 billion parameters), very fast and works at a base resolution of 512 pixels. Hands and text are unreliable and it only understands simple prompts, but the community has built thousands of fine-tunes and styles (LoRA), ControlNet and tools that still make it useful today for learning and for low-VRAM cards.

• VRAM: 4 GB or more.
• Licence: CreativeML OpenRAIL-M. Commercial use allowed, as long as you respect the licence's list of prohibited uses.
• Where to get it: model page · ComfyUI-ready version

SDXL (2023)

Stable Diffusion XL, released in July 2023, raises the base resolution to 1024 pixels and brings near-photographic quality. It's still one of the most used models today, thanks to its huge ecosystem of fine-tuned checkpoints (like DreamShaper XL or Juggernaut XL), LoRA and ControlNet, plus the fast Lightning and Turbo versions that generate in 4–8 steps. It's the best compromise if you have 8 GB of VRAM.

• VRAM: 8 GB or more.
• Licence: CreativeML OpenRAIL++-M, similar to SD 1.5's: commercial use allowed, with prohibited uses. Fine-tuned checkpoints may have their own licences: check each page.
• Where to get it: SDXL base 1.0

Stable Diffusion 3.5 (2024)

Stability AI's latest generation (October 2024), available as Large (8 billion parameters), Large Turbo (fast) and Medium (lighter). Compared to SDXL it follows prompts better and writes text better, but it was quickly overtaken by Flux and the models that came after, and its ecosystem of styles is smaller.

• VRAM: Medium about 8–10 GB; Large 12–16 GB in fp8.
• Licence: Stability AI Community License. Free for research, personal use and commercial use by companies and freelancers with less than $1 million in annual revenue; above that you need the Enterprise licence.
• Where to get it: SD 3.5 Large · SD 3.5 Medium

The Flux revolution (2024–2025)

FLUX.1 Schnell and Dev

In August 2024 Black Forest Labs, a German company founded by some of the original Stable Diffusion authors, releases FLUX.1: 12 billion parameters and a huge leap in prompt adherence, hands and text inside images. It comes in two downloadable versions: Schnell (fast, 1–4 steps) and Dev (higher quality, about 20–30 steps).

• VRAM: 12 GB in fp8, 8 GB in GGUF.
• Schnell licence: Apache 2.0, free for commercial use too.
• Dev licence: FLUX.1 [dev] Non-Commercial License. The model is for non-commercial use only, but the official page states that generated images can be used "for personal, scientific and commercial purposes" as set out in the licence. To build the model into a paid product or service you need a commercial licence from Black Forest Labs.
• Where to get them: FLUX.1 Schnell · FLUX.1 Dev

FLUX.1 Kontext and FLUX.1 Krea Dev

FLUX.1 Kontext Dev (June 2025) is the first major open model for editing images with written instructions: "change the background", "make it night", "put the same person in a kitchen", while keeping characters and style consistent. FLUX.1 Krea Dev (summer 2025), born from a collaboration with Krea, is a version of FLUX.1 Dev tuned for a more natural photographic look, less "AI-ish".

• VRAM: same as FLUX.1 Dev.
• Licence: both use the FLUX.1 [dev] Non-Commercial License, with the same rules as FLUX.1 Dev.
• Where to get them: FLUX.1 Kontext Dev · FLUX.1 Krea Dev

2025: the big open models arrive

HiDream-I1

Released in April 2025 by HiDream.ai, with 17 billion parameters and a fully permissive licence. It also uses a language model (Llama 3.1 8B) as a prompt "translator", so it handles long descriptions well. It comes in Full, Dev and Fast versions, plus a companion editing model (HiDream-E1.1). It's heavy: you'll want at least 16 GB.

• VRAM: 16–24 GB with fp8/GGUF versions.
• Licence: MIT, free for commercial use too.
• Where to get it: HiDream-I1

Chroma

A community project, trained starting from FLUX.1 Schnell and therefore under a permissive licence. People love it because it's a Flux-style model without the limits of the Dev licence, with a wide range of styles. Heads-up: it's an uncensored model, flagged as not safe for all audiences.

• VRAM: similar to FLUX.1 (12 GB in fp8).
• Licence: Apache 2.0.
• Where to get it: Chroma · Chroma1-HD (the high-resolution version, August 2025)

Qwen-Image and Qwen-Image-Edit

In August 2025 Alibaba's Qwen team releases Qwen-Image: 20 billion parameters, excellent at rendering text in images (even long text) and at following complex prompts. With Qwen-Image-Edit and its updates (versions 2509 and 2511) it becomes one of the go-to models for editing: replacing objects, changing text on signs and posters, combining several images. In December 2025 comes Qwen-Image-2512, the update to the generation model: much more realistic people (less of that "AI look"), better natural details (fur, water, foliage) and even more accurate text.

• VRAM: 24 GB in fp8; 12–16 GB with GGUF.
• Licence: Apache 2.0, free for commercial use too.
• Where to get them: Qwen-Image · Qwen-Image-2512 · Qwen-Image-Edit 2511

HunyuanImage 2.1 and 3.0 (watch the licence)

In September 2025 Tencent released two models: HunyuanImage 2.1 (17 billion parameters, generates natively at 2K) and HunyuanImage 3.0, an 80-billion-parameter giant (13 active at a time) that combines image understanding and generation, built more for servers than home PCs. Technically solid, but with a problem if you live in the EU: the Tencent Hunyuan Community License states right at the start that it "does not apply in the European Union, United Kingdom and South Korea". Its territory is defined as "the worldwide territory, excluding" those areas. In practice, from the EU there's no licence covering its use, so we recommend picking another model.

• Licence: Tencent Hunyuan Community License, excluding the EU, UK and South Korea.
• Where to get them: HunyuanImage 2.1 · HunyuanImage 3.0 (to read the licence)

Z-Image and Z-Image Turbo

The breakout hit of late 2025. Z-Image Turbo, released in November 2025 by Alibaba's Tongyi-MAI group, has "only" 6 billion parameters but produces top-tier photorealistic images in 8 steps and renders English and Chinese text well. The developer says it fits on a 16 GB card, and with the reduced versions it runs well on 12 GB too. In January 2026 Z-Image followed: the base model, better suited to LoRA training and to more varied results.

• VRAM: 12–16 GB.
• Licence: Apache 2.0, free for commercial use too.
• Where to get them: Z-Image Turbo · Z-Image

FLUX.2 Dev

In November 2025 Black Forest Labs releases FLUX.2 Dev: 32 billion parameters, generation and editing in the same model, and support for multiple reference images at once (to keep characters, products and styles consistent). It's the most powerful Flux you can download, but also very heavy.

• VRAM: 24 GB and up, even in fp8.
• Licence: FLUX Non-Commercial License: same rules as FLUX.1 Dev, non-commercial model, images usable for work too.
• Where to get it: FLUX.2 Dev

LongCat-Image (Meituan)

Released in December 2025 by Meituan, LongCat-Image has 6 billion parameters but competes with much bigger models. It focuses on photorealism and text in images in English and Chinese (put the text you want rendered in quotes in the prompt). There's also LongCat-Image-Edit for editing images. ComfyUI supports it natively.

• VRAM: the developer quotes about 17 GB with part of the model offloaded to system RAM; less with the reduced versions.
• Licence: Apache 2.0, free for commercial use too.
• Where to get it: LongCat-Image · LongCat-Image-Edit

The 2026 models

FLUX.2 Klein 4B: the Flux you can use for work

In January 2026 Black Forest Labs releases FLUX.2 Klein, the compact FLUX.2 family: generation and multi-reference editing in the same model, in just 4 steps, under a second on recent cards. For many of you the 4-billion-parameter version is the most interesting one, for two reasons: it runs on consumer cards and it has an Apache 2.0 licence. That makes it the first Flux you can use worry-free in commercial projects, products and services included. There's also a Base 4B version, non-distilled, suited to training your own LoRA.

• VRAM: about 13 GB according to the developer (RTX 3090 / 4070 and up); less with fp8 versions.
• Licence: Apache 2.0, free for commercial use, for both the model and the images.
• Where to get it: FLUX.2 Klein 4B · FLUX.2 Klein Base 4B

FLUX.2 Klein 9B

The 9-billion-parameter big sister: higher quality and the same features (generation, editing, multiple references, 4 steps), but under FLUX's non-commercial licence.

• VRAM: about 29 GB at full precision according to the developer; much less with fp8 versions.
• Licence: FLUX Non-Commercial License: non-commercial model, images usable for work too.
• Where to get it: FLUX.2 Klein 9B

Anima: the model for anime and illustration

Born from a collaboration between CircleStone Labs and Comfy Org (it's the first project backed by Comfy's open-model fund), Anima is a small model, about 2 billion parameters, specialised in anime, characters and illustration styles. The preview came out in early 2026 and version 1.0 in May 2026, in three variants: Base, Aesthetic and Turbo. It's not made for photorealism or text, but in its niche it's among the best, and it runs on low-VRAM cards too.

• VRAM: lightweight, 6–8 GB is enough.
• Licence: CircleStone Labs Non-Commercial License for the model; the official page states that generated images can be used commercially.
• Where to get it: Anima · ComfyUI tutorial

ERNIE-Image (Baidu)

Released by Baidu in spring 2026, ERNIE-Image has 8 billion parameters and is one of the best open models for text in images and layouts: posters, comics, multi-panel pages, infographics. It includes a small "Prompt Enhancer" that automatically expands short prompts. The Turbo version generates in 8 steps instead of 50.

• VRAM: the developer recommends 24 GB consumer cards; less with the reduced versions.
• Licence: Apache 2.0, free for commercial use too.
• Where to get it: ERNIE-Image · ERNIE-Image Turbo

Ideogram 4.0

Ideogram, the online service famous for text in images, released the weights of its first open model in June 2026: Ideogram 4.0, 9.3 billion parameters trained from scratch. Its strength is still text (logos, posters, covers), and it adds a unique kind of control: JSON prompts, where you set the position of elements with bounding boxes and the palette with colour codes. ComfyUI has supported it since day one.

• VRAM: roughly 12–16 GB with the fp8 or nf4 versions.
• Licence: Ideogram Non-Commercial Model Agreement. Free for research, study, personal projects and testing; Ideogram claims no rights over the images, but the free licence excludes generating images for revenue-generating products and using the images to train competing models. It also requires you to disclose that the images are AI-generated. For work you need a paid commercial licence from Ideogram. To download it you have to accept the licence on Hugging Face.
• Where to get it: Ideogram 4.0 fp8 · licence text

Boogu-Image 0.1

Released in June 2026, Boogu-Image is a 10-billion-parameter family that combines generation and editing: Base, Turbo (3–4 steps, geared towards realistic photography), Edit and Edit-Turbo for editing images (adding or removing objects, changing style). It renders English and Chinese text. ComfyUI has an official tutorial for it.

• VRAM: according to the developers it works with 12 GB by offloading part of the model to system RAM; from 24 GB up it runs without compromises. fp8 versions are available.
• Licence: Apache 2.0, free for commercial use too.
• Where to get it: Boogu-Image 0.1 Turbo · Boogu-Image 0.1 Edit · ComfyUI tutorial

Krea 2

In June 2026 Krea, known for its online platform, releases the weights of Krea 2: a 12-billion-parameter model trained from scratch, loved for its polished aesthetic and photographic look. It comes in two versions: Turbo (8 steps, for finished images straight away, up to 2048×2048) and Raw (the base model, for training your own LoRA).

• VRAM: the developer gives no official figure; roughly similar to FLUX.1, so 12–16 GB with fp8 versions.
• Licence: Krea 2 Community License. Free commercial use if the company (including affiliates) makes less than $1 million a year; above that you need the Enterprise licence. The licence also asks you to put content filters or controls in place to prevent illegal material. The images are yours.
• Where to get it: Krea 2 Turbo · Krea 2 Raw · licence

Qwen-Image 2.1

The newest model on this list (September 2026). Qwen-Image 2.1 has 7 billion parameters, so it's much lighter than the original Qwen-Image. It combines generation and editing, accepts up to 10 reference images and, best of all, generates images with a transparent background (a real alpha channel), super handy for graphics, logos, products and compositing. ComfyUI supports it natively from version 0.37, with official templates.

• VRAM: roughly 12–16 GB with the reduced versions.
• Licence: Qwen Research License, non-commercial use only. That's a big change from the original Qwen-Image (Apache 2.0): to use it for work you need to request a commercial licence from Alibaba.
• Where to get it: Qwen-Image 2.1

What about FLUX 3?

In July 2026 Black Forest Labs announced FLUX 3, a single model for images, video and audio. For now it's only available in early access through the company's own services and can't be downloaded. We'll update this page when the weights are released.

Other models worth knowing

• Lumina-Image 2.0 (January 2025, about 2.6 billion parameters) — lightweight, with an Apache 2.0 licence. Model page
• Kandinsky 5.0 Image Lite (November 2025, about 6 billion parameters) — developed in Russia by Kandinsky Lab, with an MIT licence. Model page
• NewBie Image (late 2025) — experimental anime model, with a non-commercial licence. Model page

ComfyUI also supports other, less common research models; you'll find the up-to-date list in the Comfy workflow library. Some names you'll see in the templates (like Nano Banana, Seedream, Grok Imagine or HY Image 3.5) are actually paid online services used through partner nodes, not models you can download.

Which model should you choose?

Where to download them (and how to do it safely)

• From the ComfyUI templates: the easiest way. Pick the model in the workflow menu and ComfyUI offers to download the right files into the right folders. You'll also find the workflows in the official Comfy library.
• From Hugging Face: the official pages linked in this guide, where you can also read the licence.
• From Civitai: for community fine-tunes and LoRA, with sample images and prompts. Check the licence of every file.
• Safety: stick to .safetensors and .gguf files, and avoid old .ckpt and .pt files from unknown sources, which can contain executable code.

The graphics cards we recommend

• NVIDIA RTX 3060 12 GB – the budget choice we use every day
• NVIDIA RTX 5060 Ti 16 GB – the sweet spot: 16 GB runs almost every mid-sized model in fp8
• NVIDIA RTX 5070 Ti 16 GB – same memory, much faster generations
• NVIDIA RTX 5090 32 GB – for the biggest models with no compromises
• Ready-made PC with RTX 5060 Ti 16 GB – if you'd rather not build one yourself

Frequently asked questions

What's the best image model for ComfyUI in 2026?

It depends on your VRAM and what you need. With 12–16 GB, Z-Image Turbo is the best balance of quality, speed and permissive licence. For text and posters, go with ERNIE-Image or Qwen-Image. With 24 GB or more, FLUX.2 Dev and Qwen-Image give you the best results.

Can I sell images made with Flux?

With FLUX.1 Schnell and FLUX.2 Klein 4B (Apache 2.0), yes, with no special restrictions. With the Dev versions the model licence is non-commercial, but it states that generated images can be used commercially. What you can't do is use the model to offer a paid service without a licence from Black Forest Labs. If in doubt, read the licence text or pick an Apache 2.0 model.

Why do some licences exclude Europe?

Some non-European companies, like Tencent with HunyuanImage, exclude the European Union from their licence. They don't say why; it's widely believed to be linked to the obligations of EU rules on artificial intelligence and data. If you live in the EU, it means those models aren't covered by a licence: better to use one of the many alternatives without exclusions.

Does Stable Diffusion 1.5 still make sense?

For learning, for 4–6 GB cards and for its huge library of styles, yes. For quality results today you're better off with at least SDXL or, if your VRAM allows, the newer models.

ComfyUI for beginners: a complete guide (2026)

22 Oct 2026 | AI – Artificial Intelligence, AI Tutorials, ComfyUI

Updated September 2026 · Reading time: about 15 minutes.

The images that stop you mid-scroll on Instagram, the AI videos that look like they were shot on a real film camera, the covers, the characters, the animations you see on social media every day: behind a huge share of them there's the same free program, and you can install it tonight. It's called ComfyUI.

The first time you open it, it looks like the control panel of a spaceship: boxes, colored wires, technical words everywhere. Plenty of people close it after five minutes. That's a shame, because behind that look there's a very simple idea, the same one as an assembly line. Once it clicks, ComfyUI becomes the most powerful and most open tool there is for creating with AI: no subscription, no limits, no censorship imposed by a company, and your files never leave your computer.

This guide is the handbook we wish we'd had when we started: no formulas, just the concepts you really need. By the end you'll be able to read any workflow, know which files to download and where to put them, find ready-made models and workflows, and know where to start with your first creations.

What is ComfyUI

ComfyUI is a free, open-source program you install on your computer (Windows, Mac and Linux) and use from your browser or a dedicated app. It isn't an AI model: it's the program that runs the models. Stable Diffusion, Flux, Qwen-Image, Z-Image, Wan, LTX and dozens of others are the "engines"; ComfyUI is the workshop where you mount them, connect them and make them work together.

It has become the standard for local generative AI for a simple reason: when a new model comes out, it usually works in ComfyUI from day one. Today, with the same program, you can create images, edit them, generate videos with audio, music, voices and even 3D models.

Why use it instead of an online service

• Free and unlimited: once it's installed you only pay for electricity, and you can experiment as much as you like.
• Creative freedom: no corporate filter blocking your prompt. You choose the model, and you remain responsible for respecting the law and the licences.
• Total privacy: images, prompts and reference photos stay on your computer. ComfyUI even works completely offline.
• Full control: you see and change every step, something online services simply don't allow.
• Repeatable: a saved workflow gives the same result a year from now, and you can share it as a single file.

If you want the full comparison with online services, including real electricity costs, read AI images: local or online? (in Italian)

How it works: nodes, wires and workflows

Picture an assembly line. Each station does one job: one takes the raw material, one processes it, one finishes it, one packs it. In ComfyUI each station is a node, the conveyor belts are the wires (links) and the whole line is the workflow.

• A node is a box with a name at the top. On the left it has inputs (what it receives), on the right outputs (what it produces), and in the middle the parameters you can adjust: a text, a number, a menu.
• A wire carries data from an output to an input. You can only connect data of the same type, and every type has its own color: that's why you can't really get a connection "wrong".
• The workflow is the whole set of nodes and wires. You save it as a file and reload it with one click.

The data types that travel through the wires

The basic 6-node workflow

Almost every image workflow, even the most complicated ones, is a variation of this chain. Understand this one and you understand everything else.

1. Load Checkpoint — loads the model and provides the three ingredients: MODEL, CLIP and VAE.
2. CLIP Text Encode (twice) — one node for the positive prompt (what you want) and one for the negative prompt (what you don't want).
3. Empty Latent Image — prepares the empty "canvas": width, height and how many images to generate.
4. KSampler — the heart of it all: starting from pure noise, step by step, it "cleans" it up until it matches the prompt. Here you choose the seed, the number of steps, how closely to follow the prompt (CFG) and the method (sampler and scheduler).
5. VAE Decode — turns the compressed draft into the real image.
6. Save Image — saves the result in the ComfyUI/output folder.

Why the "compressed version"? Models don't work directly on the millions of pixels in an image: it would be painfully slow. They work in a latent space, a kind of much smaller draft that holds the essence of the image. Only at the end does the VAE "develop" it into pixels, like film in a darkroom.

Run, queue and cache: how a job starts

Press Run (or Ctrl+Enter) and the workflow goes into the queue: you can line up as many as you want and go grab a coffee. ComfyUI is smart: if you run the workflow again changing only the prompt, it doesn't recompute the nodes that haven't changed (for example, it doesn't reload the model), so the second image arrives much faster than the first.

Models: what the files are and where they go

ComfyUI doesn't include any models: you download them separately. Each file type has its own folder inside ComfyUI/models, and this is where many beginners get lost. Here's the map.

File formats. Always prefer .safetensors files: they're safe because they only contain data. You'll also find fp8 versions (about half the memory, almost identical quality) and GGUF versions (even lighter, for cards with little video memory). Avoid old .ckpt and .pt files from unknown sources: they can contain executable code.

The easiest way: open a template from the workflow menu. ComfyUI offers to download the right files for you and puts them in the right folders.

Where to find models

• Hugging Face — the official "library" of artificial intelligence: this is where companies and labs publish their original models (Stability AI, Black Forest Labs, Alibaba, Baidu, Lightricks…), with the licence clearly displayed. It's the right place for base models and for versions optimized for ComfyUI.
• Civitai — the biggest community of user-made models: custom models, LoRAs for styles and characters, and workflows too. Every file comes with sample images and the prompts used, ratings and comments: it's the best place to find a specific style.
• Civitai "red" — since April 2026 Civitai has been split into two sites. civitai.com shows only content suitable for everyone, while civitai.red is the version that also shows NSFW content (short for Not Safe For Work: nudity and adult content). The reason for the split is the rules of payment processors. The account is the same and there's a single catalog; only what's visible changes. It's for adults only, and the law and the site's rules still apply there too.
• The ComfyUI Manager — from inside the program you can search for and download models and nodes without opening the browser.

Before downloading, check the licence: some models are free for commercial work too, others for personal use only, and some even exclude the EU. When in doubt, read the model's official page.

LoRA, ControlNet, upscalers: how to power up a workflow

• LoRA — a small file (often from a few MB to a few hundred) that "teaches" the model something new: a painting style, your character, a product. You add it with a node between the model loader and the rest of the chain, and you can stack several together.
• ControlNet — gives you control over composition: you draw a pose, provide a depth map or outlines, and the image follows it. It's the pros' tool for getting exactly the shot they want.
• Image-to-image and inpainting — instead of starting from an empty canvas you start from a photo or a drawing, and you can redraw just one part of the image (inpainting) while leaving the rest untouched.
• Upscalers — they enlarge the image while adding detail, so you can reach print resolutions.

Custom nodes and the Manager

Beyond the core nodes, the community has created thousands of custom nodes that add new features. You install them with the Manager, which is also a beginner's lifesaver: when you open a downloaded workflow and red nodes appear, it means some custom nodes are missing, and the Manager installs them in one click. A tip: only install popular, well-maintained nodes, because they're programs running on your computer.

Where to find ready-made workflows (and how to load them)

The trick nobody tells you at the start: images created with ComfyUI contain the workflow that generated them. Drag the PNG image into the ComfyUI window and you get the whole chain of nodes, ready to use and study. It's the fastest way to learn.

• ComfyUI Examples — the official collection of examples made by the creator of ComfyUI: image-to-image, inpainting, LoRA, ControlNet, upscaling, models like SDXL, Flux, Qwen-Image, plus video, audio and 3D. Every image contains its workflow: drag it into ComfyUI and off you go.
• The built-in templates — from the workflow menu, with a ready-configured template for almost every model. You can also find them online in the Comfy workflow library.
• Our free workflows — three workflows for Stable Diffusion 1.5, SDXL and Flux Schnell, also designed for cards with little memory (instructions in Italian). You'll find them at the bottom of the installation guide.
• Civitai — many creators also publish workflows alongside their models.

Features that make ComfyUI even easier

• Subgraph — pack a group of nodes into a single reusable "super-node", so big workflows stay tidy.
• App Mode — turns a workflow into a simple interface with just a few fields, handy to use yourself (or hand to others) without seeing the nodes.
• Loops — since version 0.36 the Start Loop and End Loop nodes let you repeat part of a workflow, for example to generate a long video in chunks.
• Partner nodes — let you use some paid online models inside ComfyUI. They're optional: if you only want to work locally, you can turn them off with the launch option --disable-api-nodes.

What you can create with ComfyUI

What computer you need

ComfyUI runs on Windows, Mac and Linux. What matters most is your graphics card's memory (VRAM): with 8 GB you can use Stable Diffusion and SDXL, with 12 GB also Flux and Z-Image in their reduced versions, with 16 GB you work comfortably with almost every image model and can start with video; for the most demanding video work you need 24 GB or more. All the details are in our guide to the best graphics cards for AI.

• NVIDIA RTX 3060 12 GB – the budget choice we use every day
• NVIDIA RTX 5060 Ti 16 GB – the sweet spot: 16 GB for almost every image model and your first videos
• NVIDIA RTX 5070 Ti 16 GB – same memory, much more speed
• NVIDIA RTX 5090 32 GB – no compromises, even for the heaviest video models
• Ready-made PC with RTX 5060 Ti 16 GB – if you'd rather not build one yourself

The path to learning ComfyUI

1. Install ComfyUI following our step-by-step guide for Windows: Desktop app or portable version, Manager and your first image.
2. Generate your first image with a template or with our free workflows, and try changing the prompt, seed and steps.
3. Pick the right models for your graphics card: start with our guide on how to generate AI images locally (in Italian).
4. Study other people's workflows: drag the ComfyUI Examples images into ComfyUI and take the workflows apart node by node.
5. Move on to advanced techniques: LoRA, ControlNet, inpainting, upscaling and then video.

We publish new tutorials, with videos and workflows, as we go: you'll find them all on the AI and ComfyUI page. Want to know when the next one is out? Subscribe to our YouTube channel (videos in Italian).

Quick glossary

Useful links

• ComfyUI on GitHub — the official project, with downloads and news.
• Official ComfyUI documentation — guides and tutorials for every model.
• ComfyUI Examples — official examples with the workflows embedded in the images.
• Comfy workflow library — the official templates, always up to date.
• Hugging Face and Civitai — for downloading models and LoRAs.
• AI and ComfyUI on Tutorial3D — all our tutorials (in Italian).

ComfyUI FAQ

Is ComfyUI hard to learn?

Less than it looks. You don't need to know how to code: you start from a ready-made workflow and just change the prompt. Once you understand the basic 6-node workflow, within a few hours you'll be able to modify other people's workflows.

Is ComfyUI free?

Yes, it's free and open source, and it works completely offline. Most models are free too; some, however, have licences that restrict commercial use.

Is ComfyUI available in other languages?

The interface is available in several languages, which you can choose in the settings, but the documentation and node names are almost always in English. The good news: once you know the basic nodes, you can follow any tutorial.

What are the red nodes when I open a workflow?

They're custom nodes you haven't installed yet. Open the Manager, use the install-missing-nodes option and restart ComfyUI.

ComfyUI or Automatic1111?

ComfyUI: it's constantly updated and supports images, video and audio. Automatic1111 is no longer developed and only works with older image models.

What's the difference between civitai.com and civitai.red?

They're the same site, with the same catalog and the same account. Since April 2026 civitai.com shows only content suitable for everyone, while civitai.red, for adults only, also shows NSFW content.

Best ComfyUI video models (2026): Wan, LTX-2.5, MiniMax H3

22 Oct 2026 | AI – Artificial Intelligence, AI Tutorials, ComfyUI

Updated September 2026 · Part of our AI and ComfyUI tutorials. For images, see our dedicated guide to AI image models in ComfyUI.

Just two years ago, making AI video on your own PC meant a few seconds of blurry, silent footage. Today, with ComfyUI, you can create high-resolution video with synced audio, dialogue and several connected shots. It's the thing that has surprised us most since we started working with local AI.

It's also the area where it's easiest to get lost: huge models, new versions every month and licences full of traps, to the point that some models can't be used in the EU at all. In this guide we put them in order, from oldest to newest: what they do, how much memory they need (with an eye on those who, like us, work with 12 GB) and what their licence really allows.

All the video models in one table

This page is an informational guide, not legal advice. Licences can change: before using a model for paid work, always read the full text on the official page linked in each section.

Video model licences at a glance

• Apache 2.0 (Wan 2.1 and 2.2, Mochi) — the most open: commercial use, modifications and integration into products are all allowed.
• LTX-2 / LTX-2.x Community License (LTX-2, 2.3, 2.5) — free commercial use if your revenue is under $10 million a year; above that you need a commercial agreement with Lightricks. No geographic restrictions. The videos you generate are yours, but the licence asks you to disclose that the content is machine-generated when you share it.
• Licences that exclude Europe (HunyuanVideo, MiniMax H3) — the licence text itself excludes the European Union: from the EU you can't use the model unless you get a separate licence from the company.

What computer you need for AI video

Video models are the heaviest of all: they generate dozens of frames at once (and the newest ones generate audio too). To work comfortably you need at least 16 GB of VRAM, ideally 24 GB or more, plus plenty of system RAM (32 GB minimum, 64 GB recommended), because ComfyUI offloads to RAM the parts of the model that don't fit on the card. With less memory you can still experiment using quantized versions (fp8, int8, GGUF), low resolutions and short clips.

To pick a card, read our guide to the best GPUs for AI. If you don't have ComfyUI yet, start with how to install ComfyUI on Windows.

The first generation (2024–2025)

LTX-Video 0.9

Lightricks' first video model (late 2024, updated in 2025) made its name with speed: short clips generated in seconds, even on cards with little memory. For many people it was the starting point for video in ComfyUI (we used it too in our tests on the RTX 3060), but today it has been overtaken by the LTX-2 family.

• Licence: Lightricks' own licence, to be read on the model page.
• Where to get it: LTX-Video

Mochi 1

Released by Genmo in October 2024, it was one of the first good-quality video models with a fully open licence. Today it's mostly of historical interest: later models look better and are lighter.

• Licence: Apache 2.0.
• Where to get it: Mochi 1

Wan 2.1 (2025)

In February 2025 Alibaba released Wan 2.1, the first open video model that could really compete with paid services: text-to-video and image-to-video, a large 14-billion-parameter version and a small 1.3-billion version for cards with little memory. A huge ecosystem of LoRAs, controls and ComfyUI workflows grew around Wan, and it's still widely used today.

• Licence: Apache 2.0, free for commercial use too.
• Where to get it: Wan 2.1 T2V 14B

Wan 2.2 (2025)

Released in July 2025, it improves motion, lighting and composition with a "mixture of experts" architecture: one model sets up the scene, another refines the details. It comes in three versions: T2V-A14B (from text), I2V-A14B (from an image) and TI2V-5B, the compact version that generates 720p video at 24 frames per second on consumer cards like the RTX 4090. It's the go-to choice if you want the most open licence. It doesn't generate audio.

• VRAM: the 5B version is designed for consumer cards (24 GB); in ComfyUI the 14B versions need quantized versions and RAM offloading.
• Licence: Apache 2.0, free for commercial use too. So far, later versions of Wan have only been offered as an online service.
• Where to get it: Wan 2.2 T2V A14B · Wan 2.2 TI2V 5B

HunyuanVideo 1.5 (watch the licence)

Tencent's video model (November 2025) has 8.3 billion parameters, generates up to 10 seconds at 480p or 720p and runs on at least 14 GB with offloading. Technically it's excellent for its quality-to-weight ratio, but the Tencent Hunyuan Community License explicitly states that it "does not apply in the European Union, the United Kingdom and South Korea". From the EU there's no licence covering its use: better to choose Wan or LTX.

• Licence: Tencent Hunyuan Community License, EU, UK and South Korea excluded.
• Where to get it: HunyuanVideo 1.5 (to read the licence)

The new generation: video with audio (2026)

The big leap of 2026 is audio generated together with the video, in the same pass: dialogue with lip sync, sound effects and ambience. No more adding the audio afterwards in another program.

LTX-2 (January 2026)

On 6 January 2026 Lightricks released LTX-2, the first production-ready open model that generates synchronized video and audio in a single pass. It has 19 billion parameters (about 14 for video and 5 for audio) and goes up to 4K at 50 frames per second, for clips up to 20 seconds long. Lightricks says it was trained on licensed data.

• Licence: LTX-2 Community License, free for commercial use too under $10 million in annual revenue.
• Where to get it: LTX-2

LTX-2.3 (March 2026)

LTX-2.3 takes the model to 22 billion parameters with a new VAE and a new latent space: sharper details, cleaner edges, better audio, more consistency from image to video and better support for the vertical 9:16 format, perfect for Shorts and Reels. As the prompt "translator" (text encoder) it uses the 12-billion-parameter Gemma model.

It's natively supported in ComfyUI, with six official templates and no extra custom nodes:

• text to video;
• image to video;
• first and last frame (FLF2V): a video that connects two images;
• image + audio: a character speaking in sync with an audio track;
• control with IC-LoRA (depth, pose, edges);
• ID-LoRA: personalized videos from a reference.

Besides the full model (dev) there's an 8-step distilled version, much faster, plus upscalers to increase resolution and smoothness.

• VRAM: it's a big model. Even in fp8 it works best with 24 GB or more and plenty of RAM; with less memory you can try the distilled version, low resolutions and short clips.
• Licence: LTX-2 Community License, free for commercial use too under $10 million in annual revenue. No geographic restrictions.
• Where to get it: LTX-2.3 · official ComfyUI tutorial

LTX-2.5 (August 2026)

On 11 August 2026 Lightricks released LTX-2.5, 22 billion parameters, which it calls an open "world model". The main new features:

• Native multi-shot: several connected shots in a single pass, with characters and settings that stay consistent from one cut to the next.
• Automatic duration: an optional module works out on its own how long the clip should be, based on the action you describe.
• Native 4K HDR and a new prompt translator (12-billion-parameter Gemma 4), better at long prompts describing characters, camera moves and lighting.
• 8-step distilled version, much faster, alongside the full model for training.

It has been supported in ComfyUI since day one, with official text-to-video and image-to-video templates and quantized int8 and NVFP4 versions (the latter for the newest NVIDIA cards).

• VRAM: like LTX-2.3, 24 GB or more is best; the int8 versions reduce the footprint.
• Licence: LTX-2.x Community License, free for commercial use too under $10 million in annual revenue, with no obligation to show the brand. The videos you generate are yours; you must disclose that they are machine-generated.
• Where to get it: LTX-2.5 · licence text

MiniMax H3 (August 2026): the most powerful, but not in the EU

MiniMax, the company behind the Hailuo service, unveiled H3 on 31 July 2026 and released its weights on 3 August. It's considered the most powerful open video model: 33 billion parameters, video with stereo audio generated in the same pass, clips from 4 to 15 seconds at 24 frames per second, a base resolution of 768p and many aspect ratios, including 9:16. It comes in two versions:

• FL2VA: from text, with optional first and last frames;
• Ref2VA: from references, up to 9 images, 3 videos and 3 audio clips.

Some parts haven't been released: the module that upscales the output to 2K and the one that rewrites prompts are only available online. ComfyUI supports it natively, and the community has already released LoRAs to speed it up (8-step versions), to control it and for video editing. The quantized INT8 version takes about 19.5 GB for the model alone.

But for us the licence is the key point. The MiniMax H3 Community License applies worldwide except in the "Excluded Territories", defined as the European Union, the United Kingdom, South Korea and the United States. Anyone in these territories is invited to contact MiniMax to obtain a licence. In the rest of the world commercial use is allowed, but above $20 million in annual revenue you need written authorization.

In practice, from the EU: without a licence requested directly from MiniMax, you can't use H3 locally. For video with audio in Europe, the open alternative is the LTX-2 family; MiniMax's online Hailuo service is a separate matter, with its own terms of use.

• Licence: MiniMax H3 Community License, EU, UK, South Korea and US excluded; above $20 million in revenue you need authorization.
• Where to get it: MiniMax H3 (to read the licence)

Which video model to choose

Practical tips for video in ComfyUI

• Start from the official templates in the Workflow menu: they download the right files and come with tested settings. You'll also find them in the Comfy library.
• Test small first: low resolution and a few seconds to find the right prompt, then raise quality and length.
• Use the distilled versions (4–8 steps) for tests: they're much faster.
• Start from an image: it often pays to create the shot first with an image model (see the image models guide) and then animate it with an image-to-video model.
• Disclose that the video is AI-generated when you publish it: some licences require it and, for realistic content, so do the platforms.

The graphics cards we recommend

• NVIDIA RTX 5060 Ti 16 GB – the sensible starting point for local AI video: 16 GB of VRAM at an accessible price.
• NVIDIA RTX 5070 Ti 16 GB – same memory, but much faster: shorter wait times on every generation.
• NVIDIA RTX 5090 32 GB – the top choice: 32 GB of VRAM, the most headroom for the big video models.
• Ready-made PC with RTX 5060 Ti 16 GB – for those who'd rather not build a computer and want to start right away.

For local AI video we recommend starting at 16 GB (RTX 5060 Ti 16 GB). The 12 GB RTX 3060 is fine for learning, but with video models it's at its limit.

Frequently asked questions

What is the best open source video model in 2026?

Technically MiniMax H3, but its licence excludes the European Union. If you work from the EU, the reference is LTX-2.5 (or LTX-2.3), with built-in audio and a free licence under $10 million in revenue. If you need the most open licence, go with Wan 2.2.

Can I use MiniMax H3 if I download it anyway?

The licence grants no usage rights in the European Union: downloading and using it from the EU isn't covered by the licence. If you need it for a project, contact MiniMax for a dedicated licence, as the licence text itself says.

How much memory do LTX-2.3 and LTX-2.5 need?

They're 22-billion-parameter models: to work comfortably you need 24 GB or more of VRAM and at least 32 GB of RAM. With quantized versions (fp8, int8), the distilled version and lower resolutions you can experiment with less memory too.

Can I use the generated videos for work?

With Wan (Apache 2.0), yes. With the LTX-2 family, yes, if your business makes less than $10 million a year, and you disclose that the content is machine-generated. With HunyuanVideo and MiniMax H3, from the EU, no, not without a dedicated licence.