← All posts
Guides

Which ComfyUI setup should I start with: pictures, video, or animating a character?

Four ComfyUI setups we got working and packed, sorted by what you want to make: what each one makes, how big it is, which rented GPUs we restored it on and how long that took.

Renest field team October 26, 2026·4 min read
Pass this on XRedditLinkedIn

Short answer: for a first picture, start with FLUX.2 Klein 4B: it is the smallest and the one we have restored most often. For video from a prompt, Wan 2.2 5B. For photo-real pictures, Z-Image-Turbo. For making a character copy the movement in a video, Wan Animate 2, which we have only restored once. Everything below is what we measured, not a ranking: we ran each of these ourselves, on RunPod machines, and we haven't compared them against anyone's benchmark.

Each setup is a starter nest: a run we got working, then packed. You take it with a free account, rent a GPU in your own account, and paste one command. See Try a starter nest for the steps, and the environments page for the current list.

I want a first picture from a prompt

FLUX.2 Klein 4B. A text-to-image setup, 12.5 GB. We restored it on an RTX 3090, 4090, 5090 and A5000. Rendering took 6 to 12 seconds on each, and the whole restore, from pasting the command to the finished test picture, took between 1 minute 37 seconds and 3 minutes 30 seconds.

A hedgehog wearing a yellow party hat, sitting on a floor covered in confetti.

We compared the pictures from the other cards with the original one side by side in FLUX.2 Klein 4B on a 3090, 4090, 5090 and A5000.

I want photo-real pictures

Z-Image-Turbo. Also text to image, 20.7 GB. We restored it on an RTX 3090, 4090 and 5090. Rendering took 9 to 19 seconds. Start to finished picture was 2 minutes 12 seconds on one 4090, 4 minutes 23 seconds on a 3090 and 5 minutes 1 second on a 5090. One more 4090 run took 30 minutes 33 seconds: nearly all of that was installing the Python packages, which took 26 minutes on that machine and between 25 and 104 seconds on the others. We don't know why.

A close-up portrait of a woman with long, wet, curly brown hair, looking into the camera, with boats and pink houses blurred behind her.

I want a short video from a prompt

Wan 2.2 TI2V 5B. Text to video, 18.2 GB. We restored it on an RTX 3090, 4090, 5090 and A5000. Rendering a short clip took 48 to 124 seconds and the whole restore 3 to 9.5 minutes. The longest wait is the download or, on the slowest run, the package install. We went through the rows in Wan 2.2 5B text-to-video on rented GPUs.

A man in a dark jacket playing an acoustic guitar by a microphone on a metro platform, a train blurring past, and two seated musicians behind him.

I want a character to copy the movement in a video

Wan Animate 2. This one takes an input image and an input video, and it is the biggest at 24.9 GB. We have restored it once, on an A40 with 45 GB of video memory, and the test render alone took about 8 minutes (472 seconds); the whole restore took 12 minutes 33 seconds. Because it is a single run, we can't tell you how it behaves on a smaller card.

Two panels: on the left a figure in a black hoodie and cargo trousers stands with arms stretched out in front of a graffiti wall; on the right a hooded figure in a black leather jacket stands in the same pose on a wet street lit by neon signs.

How the four compare, in one table

SetupMakesSizeCards we restored it onTest render
FLUX.2 Klein 4BPictures12.5 GBRTX 3090, 4090, 5090, A50006–12 s
Z-Image-TurboPictures20.7 GBRTX 3090, 4090, 50909–19 s
Wan 2.2 TI2V 5BVideo18.2 GBRTX 3090, 4090, 5090, A500048–124 s
Wan Animate 2Video from an image and a video24.9 GBA40, once472 s

Video memory on those cards: 24 GB on the 3090, 4090 and A5000, 32 GB on the 5090, 45 GB on the A40. We haven't tried any of these on a card with less.

What we haven't said

  • No ranking. We haven't compared the quality of these models with each other or with anyone else's.
  • Licences. The publishers' repositories for all four are tagged Apache-2.0. Read the licence yourself before you use one commercially.
  • Prices. We leave GPU prices out because they change by the hour.
  • "Restored" means files and dependencies came back, checked, and the workflow ran once. It doesn't mean any other workflow you load will run.
  • The times are single runs on RunPod, on the day we ran them. They will differ for you.
Pass this on XRedditLinkedIn

See a restore check itself.

Claim a starter nest — a run we packed and checked — and restore it on a machine you rent. Free, no card.