Відео

Not yet released. Once released for flux the entire workflow will need to be modified will make a new video. Sdxl requires minor changes. But the person who released the model was working on a flux release and not sdxl one.

@controlaltai Thanks for the info, so I am waiting for the release. Anyway thanks for the great worflow ❤️

Hi, The empty latent in Comfy ui is only compatible with sd1.5/ sdxl. When sd3 came out they made a node sd3 latent. The requirements for the empty latent is same at the code level for sd3 and flux. They just kept it at that.

No, I should update the node to give that option. If you want a workaround then you have to use a bunch of math nodes for multiplication and divisibility and calculate the entire formulation.

Ermm no. We need SAM segmentation. Its more accurate. Florence just created bounding boxes. The masking is done via SAM.

Noted. This is a technical tutorial aimed at helping users with ComfyUI. If the AI voice isn’t your preference, the focus is on the content and learning-feel free to skip if it’s not for you.

Hi, thank you. You can connect via email: mail @ controlaltai . com (without spaces).

HI, You need a nVidia GPU to run this. Flux Region Control uses Xformers memory attention and required a GPU. What's your GPU and system Specs?

Thanks!

SDXL only

Firstly, it should be a safetensor file and not .bin. And It goes here: ComfyUI\models\ipadapter

@@controlaltai Thanks for replying! I figured it out.

@@AndresFernandez-m9o Hi, You have to install the missing custom nodes. Check the requirement chapter of the tutorial on which nodes to install.

@@controlaltai it is working now, was using an old json, now I cant find the IPAdapter model even though I have downloaded already, where should I put it. It does not say in the video.

Goes here: ComfyUI\models\ipadapter

@@controlaltai thank you! it is working now! is it normal to have just one image in the four image outputs (the other three are red like errors)

You have to enable them. They are upscale ones. If you don’t need up scaling then just the one is enough. Normally you generate one if fine then proceed further by enabling and then que prompt.

Hi, You will require a custom workflow. Midjourney won’t solve the problem, blending will require too much post processing. IC light is only supported by sd 1.5, you cannot use flux or sdxl with IC light. We can achieve a natural blend with flux but the lighting will be pseudo. Again it would be a custom workflow. There won’t be a UA-cam tutorial for this as it’s a very specific use case.

It’s build in comfy, native node.

You need to install xformers. Regional control uses xformers for attention override. I have explain in the video. If you go to the custom nodes folder ControlAltAI nodes there is an xformers instructions there as well.

Hi, Welcome. AI moves very very fast, comfy updates breaks almost everything every now and then. All my workflows are still valid except for 2-3. For example when IP adapter changed, I updated everything. IC Light changed almost twice, and it was updated. Mostly when a user faces issues which is breaking for the entire community, I will update the workflows and post it in the members channel. Our main goal was always to maintain the workflows cause what's the point of the video then. Due to time constrains updating the workflows may take a month or two sometimes. But someone has to mention that the workflow is broke. Currently only two workflows are broken: The watermark workflow and the face Detailer. These are pending on my side. Rest of the time the workflows just require a node change since comfy updates and that specific node is not supported any more by the node dev. So if you tell me this is broken, I will fix it. There should be no issues with that, it may take some time, but it will be done.

@@controlaltai thanks for your prompt reply. I'll try to figure out how to subscribe now. Thanks for all your work.

I have joined. I see your workflows now. Thank you!

Welcome, If you need any help just write on the post or email.

@@stu75854 yeah learn and already done that in all the latest videos

@controlaltai Sorry, it's just that it makes it harder to hear you.

Not sure what the error is never heard of it. I suggest you first try this locally then read documentation on how to set up run pod

@@controlaltai The fun thing is I'm running on a pod because I don't have a pc to run on local

Unfortunately I never worked with runpod. So I can’t advise. Please look at some other tutorials on how to load checkpoints workflows etc on runpod. Once loaded. If there are workflow issues I can help you with that. Most of the clients (companies) I work work have be do the workflows locally

Only for paid channel members

Happy new year to you. Unfortunately I don’t know how to do this on run pod. Check their documentation for the same.

Thank You, So this is a very old video. Things have changed, but these values are called weights. Meaning one keyword has more weight than the other. Midjourney 6 doesn't require this. However, they have image weights when using images in prompts. For Midjourney 6, just use natural language cause AI now understands the natural detailed language better.

Hi, thanks for taking the time to watch and share your thoughts. I’m sorry the video felt confusing. My goal was to show how CosXL Edit, CosXL, and Upscale can fit into a single workflow, but I understand it might feel overwhelming if you’re new to these tools. I really appreciate your honesty-it helps me see where I can improve. I created this single, comprehensive tutorial because the workflow is unique and is my own creation. CosXL and CosXL Edit go hand in hand, so I start with basic CosXL concepts before moving on to CosXL Edit-this way, you can understand how each parameter affects the output instead of guessing. The Upscale method I use is specific to CosXL and CosXL Edit, which is why splitting it into a separate video would feel disconnected. However, you can skip the CosXL/Upscale section if you only want to focus on CosXL Edit-the video has chapters, so you can jump right to whatever interests you. This way, you have the entire advanced process in one place but can still pick and choose what you watch. If there’s a specific part you’d like me to clarify, please let me know. I want to make sure this content is as clear and useful as possible. Thanks again for your feedback!

There is clearly something else wrong here. The workflow is designed to maintain consistency in upscale. If your 1024x1024 image is clear, the upscale should be clear. The upscale is specifically designed to maintain text and not add any details than already there in 1024x1024 raw generation. If you are using any other upscale model than showcased, you will get weird results as those models were specifically chosen and tested to maintain text consistency. If some other settings have changed in the upscale process or something is not getting triggered, wrong prompt switch etc, can cause this. I cannot help you out here on UA-cam, the only way to help you is if you send me the workflow with the generated image which comes out wrong. I need to see the workflow and your image or prompt and study the entire flow to understand what is going wrong and what settings to change to get the desired output. You can email me mail @ controlaltai . com (without spaces) I will be happy to help you out over email.

@@controlaltai My 1024x1024 images aren't clear either. I suppose it's because I'm using photos as I2I source, with faces of small area, and these aren't high quality. Even if I prepare these with SUPIR beforehand, the results don't change. I can imagine, the sampler produces worse quality if the source isn't a SD image.

It is imperative for the 1024 to be clear and proper. If you are using Image to Image there are plenty of settings. Basically for image to image in the workflow is different. It is not suppose to give you the exact image. The image to image function is to generate something flux is trained on using image as a reference. But you can use image to image to change style etc, but it any case it wont be the same, it will change and the objective is to generate something new from your reference image no matter what resolution. Now Image to Image will generate a proper flux 1024 output depending on the prompt. Your prompt has to be detailed. Once you get a proper 1024x1024, upscale will maintain that. If you want to enhance the 1024, then thats a different workflow. If I can see what Image to Image you are trying to achieve, I can help with the settings and how to go about it. Cause its a bit tricky, (too many minor settings to play around with).

Hi, Check the post processing section at the end. If its dark and poorly lit, bypass post processing. That should have no added leveling or contrast. You can then use those post processing nodes to increase brightness color correction etc. Preserve Details Noise Plus Blend has zero effect on the gloomy, its most likely the auto correct lighting post processing node.

@@controlaltai It's the "Color Match", spoiling the highlights. Thanks!

@@controlaltai The Tiles CN is making the final output gloomy, too. I don't use it any more.

The ControlNet is extremely difficult and tricky to use. May I know for what purpose are you using the tile controlnet for, cause what we used to use tile for SDXL doesn't actually apply or work with flux, its a different architecture. Tile can be used for say enhancing the image, generating 3d out of 2d, and most of the time the tile has to be daisy chained with some thing else. Also the the strengths of CN in flux are not like SDXL. The Steps % play a more critical role, that is, they are more sensitive to CN than CN strength. Which is very different than how CN works on SDXL. For example to get a pose, I prefer to use Depth rather than DW pose or if there are hands, then i use both. IN SDXL we would just use dw pose.

There is no multi regional Lora support with this or any other current solution at the moment. You can apply Lora it will apply to the whole region.

Hi, expand the post the workflow link is given....Here it is again Ollama might not work on runpod. I have no experience with that. Instead of Ollama you can install Gemini Node or any Cluade AI node which uses API, API based LLM would easily work on runpod. For the instructions use the same I have given in the workflow. All you have to do is replace Ollama with your own API Custom node (whatever custom node you choose) and ensure the connections are all correct. Workflows: drive.google.com/file/d/1WH5Exmzij-shWnQ7MQOagE_jTTL8pEtx/view?usp=sharing Workflow Images: drive.google.com/file/d/1z_9my1bzxNEEArGWsFzHWlINl60BCtdY/view?usp=sharing

@@controlaltai Thank you very much. Ill try

Welcome, Not possible in the current pipeline. Comfy recently added flux attention mask in the flux model natively. I am working on a new pipeline called "Flux Attention Mask". That should allow to add multi lora train character in image. The current pipeline will only support lora's which apply to the entire region, not multi region. It will take some time to make that pipeline as you can image its a bit complicated. However, a lot of progress has been made regarding that. Once everything is ready and stable, first the code will be released, and shortly after the video tutorial.

@@controlaltai thanks hopefully can be done soon :)

@@Ozstudiosio Yup I am at the sampling stage. But I am not sure if flux itself supports regional lora, whatever research done so far suggest it doesn’t however there might be image to image support or something after using single Lora, I will know more in some days. If the model weight supports it ,it’s straightforward.

@@controlaltai thanks but there any other method to allow me compose multi train character in scene?

Not in Flux. In painting can be a solution. But I haven't explored yet. The last option is face swap if its only face lora. Regional Conditioning is the best bet, if its supported. Instant X does it with PULID regional faces maintained by PULID, but the PULID code is not migrated to comfy. Someone else is working on that and has ported the Instant Research to Comfy UI (region only without multi lora, as of now). The instant x team said not possible to have multi lora in their system as well.

Slow down the playback. Some people find it fast and some complain its too slow. I can't adjust to meet everyone's speed. But you can with slow and fast playback on you tube. Sorry if the tutorial is too fast, I have slowed down with a consistent speed in the latest videos (learning from feedback).

There is some issue with the models. Use SD 1.5 checkpoint and not SDXL or Flux (they are not compatible). If you still get that error, then check if the ic_light models are correct and the IP adapter models are correct.

I already checked and use everything version SD 1.5, but it still have the same problem in the KSampler.

You can email the workflow to mail @ ControlAltAI . Com (without spaces). I cannot help further without looking at the workflow. Make sure to include your input product photo and reference image you are using.

You have to update comfy. then update the custom node. close everything and then restart. Check for fetch updates again to ensure you are on the latest version of the node.

@ thanks I’ll try again

That is a different GitHub repository. Install ControlAltAI nodes from Comfy Manager. github.com/gseth/ControlAltAI-Nodes