Ernie Image Turbo | LOW VRAM Workflow

Hinzugefügt: 2026-05-04

[00:00:08]All right, YouTube. What is good? It's your boy Rebel back with another video and today we are taking a look at the Ernie image turbo guff. So Ernie image turbo is an open text to image generation model developed by the Ernie image team at Badu. It is the distilled release of Ernie image built on the same single stream diffusion transformer family and designed for fast generation and strong fidelity in only eight inference steps. So essentially this model can generate highfidelity images in about eight steps similar to Z image turbo. Um, to get started, you're going to need to head to the link in the description and go to my Civid AI page and download the workflow and the model files and then you can come into Comfy and drag the workflow in and it will look something like this. So, starting on our left, we have the unit loader hosting our Guff. We have the clip node hosting our ministral 33B model um, which is running on the flux architecture. And then I have implemented the flux 2 small decoder just to save VRAM. Um you are welcome to use the full flux 2 DAE if you would like to. Um that's up to you. I think that the small decoder works just as fine. Down here we have our empty flux 2 latent which is where you will handle your resolution size. We have our positive prompt node and then we have a conditioning zero out for the negative.

[00:01:32]In our K sampler we have our seed, our eight steps, our CFG of one. You can play around with samplers and schedulers and a D noiseis of 1.0. So, of course, I'm going to have some examples for you guys today. So, we will take a look at those. So, a few of these were generated to save time at 512 x 512, but you'll get the picture from the image itself.

[00:01:51]So, this is a man wearing a black Nike sweatsuit holding a sign reading Ernie image turbo guff. Um, pretty straightforward. I do this for all my models. Um, here we have a car in New York City. Here we have a German Shepherd on a beach at sunset. Here we have a billboard in like Las Vegas. Um, this was just to test the text ability.

[00:02:14]Here we have a portrait of a man in glasses. Uh, here we have a woman at like a night rave.

[00:02:22]Here we have a man in some kind of like tactical suit. Here we have a woman in a dress in a kitchen. Here we have a man taking a walk in a forest. Uh here we have a woman in a headdress um in a field of flowers. Very nice image. Uh here we have a man in a coffee shop.

[00:02:42]Here we have a business woman in a pants suit um outside maybe downtown. Uh here I wanted to overload and stress the prompt out. So the prompt is extremely long- winded and the output it gave us.

[00:02:57]I mean it's pretty much adhered to everything that it's asked for. Um, there's a few things missing, but it's it's pretty close. Here we have a Asian woman in a coffee shop drinking a maybe tea. Try tea. There seems to be a very heavy Asian influence in these images.

[00:03:18]Like 90% of the subjects that I generated were Asian. So, keep that in mind. Here we have a futuristic woman with some silver jewelry in what appears to be like a high-tech facility or a spaceship maybe. Here we have a man in a zip-up jacket in maybe downtown Tokyo at night. Here we have a old school PC monitor and keyboard with some code and an application running. Here we have an old man sitting on a porch. Here we have a decal for Rebel AI. I thought this was pretty good. And here we have an abstract, kind of like a hologram, I guess you would say, of a brain in the woods. Just a random Gemini prompt. Now, as far as generation time, um, most of these images generated in about a minute. All right, guys. That's all I'm going to have for you today. Just wanted to get you this model and this workflow so you guys could test it out. Um, I know that it has been a little bit stale lately with image models, so it's nice to get something new to be able to test out. And the fact that this runs essentially at the same speed as Z image, you know, you're able to get a couple different looks. So, if you enjoyed what you saw today, please consider subscribing to the channel and leaving me a comment on the video to let me know how I did. You can also like the video. It genuinely helps the channel tremendously. So go ahead and download this workflow, download these models, and enjoy Ernie Image Turbo, guys. All right, thank you. Have a nice day.