Run AsymFLUX.2-Klein (AsymFlow) for FREE in Google Colab | ComfyUI

Ajouté : 2026-05-24

[00:00:00]Hello everyone, it's AI with Chucky and welcome back to the channel. Today we'll be looking at Asimalux 2 Klein, an advanced image generation pipeline which you can use right inside Google Collab for free. Now if you've been keeping up with the rapid pace of open-source artificial intelligence, you've likely heard about the release of Asim Flow developed by researchers at Stanford University and Adobe. Asymmetric flow modeling or asimflow is a fascinating approach that fine-tunes latent diffusion models directly into pixel space flow models. This allows for the generation of incredibly detailed photorealistic images without requiring a traditional VAE decoder. Instead of struggling to run these memory heavy models locally on limited hardware, we are going to set up our generation ecosystem using a free Google Collab notebook. To start, let's take a look at the official GitHub repository for LCON Lab, which contains the official codebase for Asim Flow, PyFlow, and GM Flow. As we scroll down through the documentation, we can see that as Flow fine-tunes Flux.2 Klein into a pixel space flow model. By shifting the noise prediction process to a low rank subspace, it achieves state-of-the-art results for pixel space image generation. Now, let's switch over to Google Collab. We'll start with a clean slate. I have a custom Jupyter notebook ready for us called asflux.pyb.

[00:01:31]I'm simply going to drag this file from my local downloads folder and drop it into the upload section of the collab workspace. It takes just a moment to upload and once it's loaded, we will see the visual framework of our environment.

[00:01:46]The notebook we are using is titled Asam Flux Botline Pixel Space Confu. Before we click anything, the absolute first step we must perform is checking our runtime settings. Let's navigate to the top right corner of the collab interface. Click on the connect drop-own menu and select change runtime type.

[00:02:06]Under the hardware accelerator options, make sure T4 GPU is selected. Once you've confirmed the T4 GPU is selected, click on save. Now, let's hover over to our very first code cell titled build confview ecosystem and custom extensions. We'll click the play button on the left of the cell to initialize our environment. While this first cell runs, let's talk about what's happening behind the scenes. This starts by updating system repositories and installing dependencies. It then clones Confi UI repository to our directory.

[00:02:40]Next, the script automatically installs the necessary custom extensions.

[00:02:46]This includes the comfy manager, which helps us manage our custom nodes and the core dependencies of Lacon Lab. As you can see from the active terminal output scrolling by, packages are being unpacked, built, and compiled. This part of the setup takes roughly 2 to 3 minutes. We just need to sit tight and let it complete. When it finishes, you'll see a clean line at the bottom of the output declaring environment compiled. With our workspace compiled, we move on to step three, downloading our models and workflows. We'll scroll down to the second cell titled rapid multi-threaded model and workflow acquisition, and click the play button.

[00:03:28]This cell is preconfigured to fetch the exact model weights we need directly from hugging face and other verified repositories because Google collab has extremely fast internet speeds and combined with Arya 2C we installed in cell one 2 minutes should be enough dot first the script downloads our predefined asimflow comfy UEIE workflow file so we don't have to go out there looking for workflow file to run the model later on next it pulls the main model checkpoint points. Additionally, we need a text encoder so the model can understand our prompts. For this, we are downloading the compact Quen 3 text encoder named Q A38BFP8MIX sephotensors by using lightweight FP8 or 8-bit precision models. We drastically reduce our VRAMm footprint while keeping the quality virtually indistinguishable from full precision formats. This is what allows a standard T4 GPU with 15 GB of VRAM to successfully handle such a capable generation pipeline. Once the progress bars finish updating and the files are successfully saved to their respective directories inside Confuey, the cell will complete and display assets ready. Now that our environment is built and our models are in place, we need a way to access the Confui graphical interface running on Google's remote server. This is where our third cell expose secure tunnel and launch confuite server comes in. Let's click the play button to run it. This script boots up our Confue server in the background and initializes a secure tunnel using a service called Pingi. The tunnel acts as a bridge, generating a public URL that securely redirects traffic from our browser directly to the port where Confuei is running inside our Google Collab container. In a matter of seconds, you'll see a green highlighted link appear in the output log. This link will take about a minute or two to get activated. So, if you click on it earlier, it would take you to an error page unless you give it a minute or refresh. So let's click on this generated URL. Because we are accessing a temporary public tunnel, Pingi will show a brief caution screen. This is completely normal for free tunnel services. We just need to click the enter site button. Once we click that, the Confu landing page will load, presenting us with our blank canvas. To load our workflow, let's navigate to the left side menu in Confuee. Click on the workflows tab to open the browser panel.

[00:06:04]Here you will see our pre-downloaded Asenflux 2 workflow file readily available. Let's click on it to load the node structure onto our workspace. Now let's zoom out and look at the layout of this generation pipeline. This is a beautiful highly customized node network. Let's trace the signal flow together. On the far left, we have our model loaders. The load asflow model node is loading our base model flux 2 clinbase billion parameter FP8 safe tensor file along with three asim flux 2 client adapter safe tensors including Z image beneath it. The load clip node handles our Quen text encoder. We also have the Oaklab color encoder node which is critical because ASMF Lux operates directly in the Oaklab pixel space rather than a standard latent space. In the middle, we have our clip text encode nodes for our positive and negative prompts. Below that, the image size node determines our output dimensions. I'll leave it 1024x440 pixels. On the right, we have our scheduling and sampling block. This features the asimlux 2uler and k sampler select nodes, which feed into the sampler custom advanced node to compute our image generation steps. Finally, everything ends at the VA decode and save image nodes on the far right.

[00:07:28]Before we start our first generation, let's tweak a helpful setting. If we go to the main settings menu in Confy UEIE, we can search for the preview settings.

[00:07:38]By default, the preview method might be set to default or none. I'm going to change this to auto under the Confy execution settings. This will allow us to see a live lowresolution preview of our image resolving step by step as the sampler runs. This will slow down your generation time, so you can always switch it back off if you want to. Now, let's look at our positive prompt. The default prompt reads, "Fashion photography of a beautiful young woman with floral tattoo on her cheek of the curly word as flow. Dramatic lighting, masterpiece, raw photo, Kodak film photography, detailed ring. In the negative prompt box, we have standard quality filtering tokens, lowquality, worst quality, deformed, bad anatomy, unclear text, bad skin. I'm also going to add incorrect fingers to help guide the model on hand details. Everything looks correct. Let's head over to the top menu and click the run button to cue our prompt. While the interface waits, let's check back on our Google Collab terminal tab. Here we can see the backend logs showing the model loading into memory. The text encoder and the main 9B transformer are being allocated.

[00:08:52]Because this is the highly optimized 8-bit client model running alongside asim flow, it fits cleanly within our T4 GPU's VRAMm limit. Once the model initialization is complete, the generation steps will begin. Returning to our Confuey tab, we can see the generation process has started. Thanks to our live preview setting, we can watch a hazy outline gradually form and gain detail. The sampler is running for 32 steps because we are running this on a free Google Collab T4 GPU. You will notice that the generation process is extremely slow. It takes a significant amount of time to compute each step, so you'll need a bit of patience here, except if you decide to go for Google Collab Pro GPUs. And there we go. The generation is complete. Let's right click on our output image and select open image to view it in full resolution. The model is exceptionally good with skin texture and photo realism. The lighting on the face is beautifully balanced, catching the highlights perfectly. Most importantly, look at the rose tattoo on her cheek.

[00:09:59]The text spelling of Ace and Flow is very clean with the letters holding their shape accurately. Bypassing the traditional VAE really helps preserve textual details. However, there is a catch. If you look closely at her hand and the fingers resting near her chin, you will see some severe anatomical distortions. The fingers are jumbled, merged, and look incredibly unnatural.

[00:10:23]While this model excels at photo realism, it is very poor with anatomy, which means you have to be extra and a bit aggressive in refining your negative prompts to fix these artifacts.

[00:10:35]Let's put the model to another test. A common pain point for many texttoage models is attribute binding, specifically keeping colors and adjectives attached to the correct objects without bleeding into other areas of the image. I have a test prompt ready inside my Gemini browser window.

[00:10:54]Let's copy this prompt. A portrait of a young woman with bright crimson hair tied in a thick braid wearing a vivid emerald green silk blouse and a bright mustard yellow scarf. She has one striking blue eye and one warm brown eye. Heterocchromia. Crisp studio lighting against a pure gray background.

[00:11:14]This is a very demanding prompt with multiple distinct color assignments.

[00:11:18]crimson hair, green blouse, yellow scarf, and two different eye colors.

[00:11:23]Let's head back to comfy, paste this into our positive prompt box, and hit the run button. Once again, we can watch the live preview as the pixels resolve directly. Let's open the finished image to inspect the details. The attribute binding for the clothing and hair is handled incredibly well. Her hair is a rich crimson red. Her blouse is clearly emerald green and her scarf is a vibrant mustard yellow. The colors are perfectly separated with absolutely zero color bleeding. However, the model completely failed the heterocchromia test. Both of her eyes are rendered as a striking blue, completely ignoring our prompt for one brown eye. While it manages broad color assignments well, it clearly struggles with nuanced asymmetrical facial features like mismatched eyes.

[00:12:13]For our final and most complex test, we will evaluate spatial awareness and reflection consistency. Rendering reflections in mirrors is a classic failure mode for AI models as they often struggle to differentiate between the 3D space of the main subject and the 2D plane of the mirror reflection. Let's copy our next prompt from Gemini. A woman standing directly in front of a perfectly clean vanity mirror. The left side of her face is covered in intricate neon blue cybernetic tattoos while the right side is completely clear. The camera angle captures her side profile and her reflection in the mirror simultaneously. The reflection perfectly matches the asymmetrical tattoos. This requires the model to understand geometry, perspective, and strict physical constraints. Let's paste this prompt into Confuey and click run. The sampler gets to work, and as the image resolves, we can see the dual profiles taking shape. Let's open the final highresolution output. Unfortunately, the spatial layout and reflection consistency are a complete failure here.

[00:13:23]While the neon blue cybernetic patterns are sharp and luminous, the model entirely misunderstood the geometry of a mirror reflection. The placement of the glowing lines on the foreground subject does not anatomically match the reflection we see in the mirror. The perspective is heavily distorted.

[00:13:43]Placing the tattoos on the wrong side and confusing the angles entirely. It's clear that rendering accurate logical reflections is still a major weakness for this model. Running as Flux 2 Klein directly in pixel space yields some truly impressive skin textures and localized details, but as we've seen, it requires heavy prompting to fix anatomy and struggles with complex spatial logic. To wrap it up, let's go back to our collab tab. Click on the top right menu options and select disconnect and delete runtime. Click yes to confirm.

[00:14:18]This completely terminates our session and frees up the server resources. If you enjoyed this tutorial and want to see more step-by-step guides on running cuttingedge AI models for free, make sure to like this video, subscribe to the channel, and hit the notification bell. Let me know in the comments what you want us to run next on Google Collab. Thank you so much for watching, and I'll see you in the next one.