Model × Task Overview: Pick the Right Model Before You Start
💡 Goal: Lock in the right model for the task at hand within 10 minutes — whether your VRAM is enough, whether it can be used commercially, and which tutorial to start from, all answered on this page; 15 ComfyUI workflow tutorials are on standby, categorized by image and video.
People often ask which models ComfyUI supports — the answer depends on your task and your graphics card. ComfyUI itself is just an "engine"; the models do the real work: choose correctly, and you can generate images in 8 steps or run video with 8GB of VRAM; choose wrong, and you only find out it won't run after downloading 20-plus GB of weights. So spend ten minutes first finding your fit.
🎯 Stage Goals
- Lock in candidate models within 30 seconds of receiving a task: text-to-image, image editing, transparent backgrounds, product mockups, text-to-video, first/last frame… use the matrix table to find the model and its corresponding tutorial
- Weigh your VRAM before downloading: check the VRAM references in the matrix and stop blindly downloading large weights
- Distinguish the boundary between "generation" and "editing" models: for photorealistic text-to-image, go with base models; for background replacement and text editing, go with Edit-type models; for inpainting and outpainting, go with Fill
- Run at least one new model workflow end to end: load the official workflow from the template library, put models in the right directories, and follow the tutorial's parameters to produce your first image or video clip
⏳ Estimated Time
- Completing this stage in full: 15 lessons totaling about 204 minutes (about 3.5 hours), with about 140–265 minutes of buffer
- Excludes model download time: many models in this stage are large weights at the 12B–20B level; the full FLUX.1 version is about 23GB, extra time depends on your network speed
- Skipping around as needed saves time:
- Just want to generate images ASAP: head straight to 2.1 Z-Image-Turbo, 12 minutes
- Have 8GB VRAM and want to try video: head straight to 2.10 Wan2.2 5B, 15 minutes
- Mainly doing image editing and e-commerce images: 2.3 Qwen-Image-Edit → 2.4 Qwen-Image-2.1, about 30 minutes, then continue with E-commerce Images in Practice
- Want clips with sound: 2.14 LTX-2/2.3, 15 minutes
🚀 Before You Start
- Completed 1.5 First Image Generation: able to load models, write prompts, and run with
Ctrl + Enter - Know where model files go: have a clear idea of the roles of the
diffusion_models,text_encoders, andvaesubdirectories (if you forget, review 1.4 Where to Put Models) - Update ComfyUI to the latest version: the models in this stage are all very new; the FLUX.3 Klein template requires 0.41.0 or above
- Check your disk and VRAM: large model files are big (the full FLUX.1 version is about 23GB); your VRAM size determines which lesson to learn first (see the matrix below)
- A cup of water ☕ — while models download, it's a good time to wet your whistle
💡 The template library is the main entry point for this stage
Every lesson starts by loading the official workflow from the Template Library (the Templates icon in the left sidebar); when a model is missing, the corresponding node turns red — just click Download one by one in the Errors tab on the right. For details of the whole process, see 4.1 Template Library.
📋 Choose a Model by Task
This stage covers five model families with complete official tutorials: Z-Image, Qwen-Image, FLUX, Wan2.2, and LTX. The two tables below are ComfyUI model recommendation and comparison quick-reference tables sorted by task; VRAM requirements and step counts are taken from the official tutorials — just look them up when choosing.
Image Tasks
| Your Task | Recommended Model | Key Facts (VRAM / Steps / License) | Tutorial |
|---|---|---|---|
| Fast image generation, limited VRAM | Z-Image-Turbo | Alibaba Tongyi 6B distilled version, generates images in 8 steps; runs with 16GB VRAM; accurate Chinese and English text rendering | 2.1 |
| Chinese text in the image | Qwen-Image (including 2512) | Alibaba Qwen team 20B; strong Chinese and English text rendering; distilled version 15 steps, Lightning LoRA 8 steps | 2.2 |
| One-sentence photo editing: swap backgrounds, change text, replace materials | Qwen-Image-Edit (2509/2511) | instruction-based editing; 2511 supports material replacement and character consistency | 2.3 |
| Generation + editing in one, transparent background, 2K | Qwen-Image-2.1 | single model with dual use; native 2048×2048; alpha transparent background; 16 reference image slots | 2.4 |
| Photorealistic text-to-image (English prompts) | FLUX.1 dev / schnell | 12B, more realistic and diverse images; schnell in 4 steps, Apache 2.0; dev is non-commercial | 2.5 |
| Character consistency, multi-round editing | FLUX.1 Kontext | 12B iterative editing, maintains character and style across multiple rounds; English prompts only | 2.6 |
| Local inpainting, outpainting | FLUX.1 Fill | dedicated to inpaint / outpaint; officially said to be second only to the commercial Fill [pro] | 2.7 |
| Multi-reference consistency, product mockups | FLUX.2 Dev | up to 4MP photo-grade; up to 10 reference images; hex brand colors | 2.8 |
| Fast image generation on low VRAM, previews | FLUX.2 Klein | 4B; distilled, 8.4GB VRAM, about 1.2 seconds per image (5090) | 2.8 |
| Run the latest generation of FLUX locally | FLUX.3 Klein | text-to-image + editing with the same set of weights; distilled in 4–16 steps; FLUX Kommunity License | 2.9 |
Video Tasks
| Your Task | Recommended Model | Key Facts (VRAM / Steps / License) | Tutorial |
|---|---|---|---|
| 8GB VRAM for video | Wan2.2 5B | single model for text-to-video + image-to-video; runs on 8GB VRAM with native offloading | 2.10 |
| High-quality text-to-video | Wan2.2 14B T2V | MoE with high/low-noise dual experts; Apache 2.0, commercial use allowed | 2.11 |
| Bring photos to life | Wan2.2 14B I2V | upload a first-frame image to generate a dynamic video | 2.12 |
| Control the start and end of a video | Wan2.2 FLF2V | given the first and last frames, the middle is auto-completed | 2.13 |
| Videos with sound | LTX-2/2.3 | 19B audio-video in one, generates synchronized video + audio in a single pass; distilled in 8 steps (CFG=1) | 2.14 |
| Multi-shot narratives, 4K HDR | LTX-2.5 | native multishot; 4K HDR up to 50 FPS; access must be requested first for the HF repo | 2.15 |
⚠️ Check the license before commercial use
schnell and Wan2.2 are Apache 2.0 and can be used commercially; the FLUX.1 dev series (including Kontext and Fill) is under a non-commercial license; FLUX.3 Klein uses the FLUX Kommunity License. For other models, refer to the license on the download page.
💡 If you can't decide, start with the lowest-barrier option
When VRAM is tight: for images, try FLUX.2 Klein (distilled, 8.4GB) or Z-Image-Turbo (16GB) first; for video, try Wan2.2 5B (8GB) first. Once you get one running, the workflow structure of the other models is largely similar—switching models is much easier than learning a new tool.
📚 Courses for This Stage
| Course | Content | Duration |
|---|---|---|
| 2.1 Z-Image-Turbo text-to-image | 8 steps, 16GB VRAM homegrown speedster; Fun Union ControlNet and 2K upscale | 12 minutes |
| 2.2 Qwen-Image text-to-image | 20B base model writes Chinese text well; distillation / Lightning acceleration; 2512 photography upgrade (downloadable on ModelScope) | 12 minutes |
| 2.3 Qwen-Image-Edit instruction-based editing | change backgrounds and edit text with a single sentence; 2509 multi-image consistency, 2511 material replacement | 15 minutes |
| 2.4 Qwen-Image-2.1 | generation + editing in one; native 2K, transparent background, 16 reference images | 15 minutes |
| 2.5 FLUX.1 text-to-image | local deployment of dev and schnell; full and fp8 installation methods | 12 minutes |
| 2.6 FLUX.1 Kontext | character-consistent multi-round iterative editing; chain edits by pulling images from the output directory | 12 minutes |
| 2.7 FLUX.1 Fill | inpainting and outpainting | 12 minutes |
| 2.8 FLUX.2 | Dev multi-reference consistency and the Klein lightweight version; product mockup workflow | 15 minutes |
| 2.9 FLUX.3 Klein | the latest open-source FLUX; text-to-image + editing with the same weights, distilled 4–16 steps | 12 minutes |
| 2.10 Wan2.2 5B | run text-to-video and image-to-video with 8GB VRAM | 15 minutes |
| 2.11 Wan2.2 14B text-to-video | dual high/low-noise expert models for cinematic results | 15 minutes |
| 2.12 Wan2.2 14B image-to-video | bring photos to life; note that I2V uses fp16 weights | 15 minutes |
| 2.13 Wan2.2 first-and-last-frame-to-video | control the beginning and the end, with the middle auto-completed | 12 minutes |
| 2.14 LTX-2/2.3 | generate video with sound in one pass; six official workflows | 15 minutes |
| 2.15 LTX-2.5 | multi-shot narrative, 4K HDR and DFR; the HF repo requires requesting access first | 15 minutes |
🎒 What you'll gain
- A model-selection map of "task → model → tutorial": when new models come out later, you'll be able to judge them yourself across the dimensions of VRAM, steps, license, and task
- Baseline awareness before downloading: check the VRAM reference and the license first (Apache 2.0 / non-commercial / Kommunity), and never waste hours waiting on a download again
- 15 official template walkthroughs matching the ComfyUI workflow tutorials: template loading, model placement, parameter tuning, all the way to producing images and videos
- Entry points to two advanced tracks: images lead to e-commerce image practice, control, LoRA and quantization lead to advanced techniques