Follow our Work on Youtube Channel:
www.youtube.com/@AiMotionStudio
Welcome to our latest tutorial on the Flux.1 Dev Controlnet Upscaler Workflow Update, featuring Flux Lora and Florence2. In this video, we walk you through everything you need to know about improving your upscaling workflows using these powerful AI tools. Whether you're an experienced user or new to AI upscaling, this video will help you optimize your workflow and resolve common issues, especially for those using lower-end GPUs.
What You’ll Learn:
In this comprehensive walkthrough, we demonstrate how to make the most of the Pro Post Film Grain node. If you like your images to have a natural, grainy look as though taken by a non-professional camera, this feature is for you! But if you prefer a cleaner, sharper image, we also show you how to bypass the grain node to fully appreciate the crispness that the upscaling process can deliver.
For users experiencing 'out of memory' errors on low-end GPUs, we provide an effective solution: bypassing both the Apply Controlnet and Load Controlnet Model nodes. Don’t worry—by relying solely on the Flux Lora model, you’ll still achieve great results with minimal memory usage. This simple trick can save you from frustrating crashes and help keep your workflow smooth and efficient.
We also dive into troubleshooting tips for when some nodes in the workflow fail to work as expected. We guide you step-by-step on how to install missing custom nodes using the ComfyUI Manager. By doing this, you can resolve any compatibility issues and ensure your software is up-to-date. It’s quick, easy, and essential for maintaining a seamless experience.
Why This Workflow Update Matters:
This update to the Flux.1 Dev Controlnet Upscaler Workflow is a game-changer, especially for creators looking for flexibility in handling large image files and AI-driven upscaling processes. Whether you are a digital artist, content creator, or AI enthusiast, this workflow ensures that you can upscale with precision and reliability. 😊🚀
Description
Workflow Update with the New Flux.1 Dev Controlnet Upscaler
FAQ
Comments (27)
I get blurry images, what could be the problem?
try to use reduce the controlnet weight to 0.3
Try this also, bypass ControlNet to see if it works. If it does, bypass LoRA next to isolate where the issue is. If LoRA works, try reducing the weight of ControlNet, as there might be a conflict with LoRA's weight. This should help identify the source of the problem. Let me know what you find.
@AIMotionStudio Tried all the options, the result is the same as it was.
Same problem, blurry images, worse than the original. ControlNet is off, tried turning off Lora, doesn't help.
After tests, it was noticed that changing upscale_method to lanczos helps a lot to make the image much sharper. But some images still turn out very blurry, while others can turn out very well. It is not clear what it depends on. ControlNet is off.
same here. The image is cleaned up but seems unfocused
@BentonTramell try to bypass the controlnet and it will work fine!
hey, its super slow. i have rtx4080 with 12gb vram. and it took 10 minutes and still no result.
can you help me ? thanks
First, bypass ControlNet to see if it works. If it does, bypass LoRA next to isolate where the issue is. If LoRA works, try reducing the weight of ControlNet, as there might be a conflict with LoRA's weight. This should help identify the source of the problem. Let me know what you find.
I have tried 5 workflows attemting to utilize this new model and only 1 worked so far: CG Top Tips. Even then, it worked better as a refiner than an upscaler, which was too slow on a 3090. I expect a fix, because development is going so fast. It is pretty cool to learn Flux in almost real time. Big Kudos to the ComfyUI community.
TypeError: apply_controlnet_cond_func_advanced() got an unexpected keyword argument 'vae'
I am getting this error. Anyone knows hot to fix it?
Hi there, is there a way to add prompt manually? The way it is by default dosent generate a correct prompt for certain issues. thanks
yes you can bypass the florence2 and then add prompt manually by adding prompt encoder text to replace the one connected though florence2.
@AIMotionStudio Great, I'm going to try it. And by the way, I have a question... is it normal for the process to take so long? I have 12 RAM GPU and 64 RAM CPU and it took me maybe an hour, I haven't counted it but it took a long time. I was thinking of putting the Flux Lora Turbo to do it in 8 steps, but I don't see the way. The truth is that I don't understand it well, because it says 6 steps but it's 10 turns. If you can give me some clue I would appreciate it... thanks for answering.
@tupu yes it can take a while for 12 GB VRAM but you can change the Half tile + intersections to Band Pass and also bypass the Controlnet to reduce the time it takes to upscale by half and this also get the job done. leave everyother settings as default
@AIMotionStudio Thanks, really appreciate!! Good job, very nice presentation!!
UltimateSDUpscale
.to() does not accept copy argument
An error appears. Is there a solution?
try to update your ComfyUI to the latest version.
Also try the version 1.0
I have the same problem
Guys, if your output picture is pixelated or unfocused, just change these factors:
- Change Upscale Method to lanczos
- Decrease ControlNet Strength
- Increase Ultimate SD Upscale Denoise
- Decrease CR Lora Stack weight
this worked for me.
Ultimate SD Upscaler doesn't work and Confui informes: UltimateSDUpscale mat1 and mat2 shapes cannot be multiplied (10126x16 and 64x3072) What shall I do?
try thr Version 1 see if that works, if not then you should update confyUI!
you're probably using the wrong version of pytorch and it won't fork with the nf4 model or make sure that bitsandbytes is properly installed
try going into the ComfyUI\custom_nodes\ComfyUI_bnb_nf4_fp4_Loaders folder. open cmd
type in the command pip install -r requirements.txt
I'm skeptical about the result, in my tests it tends to smooth out details more than enhance them. It might not be well suited for illustrations
this used to work perfectly well, but i think the checkpoint loader is corrupted, i get the dimension error and i cant fix it what ever i do