Відео

Download my workflows, update your ComfyUI.

Thanks a lot! I've already updated the description for this video and added a new workflow for Flux Fill. Regarding Depth and Canny I'm not sure as we already have several good solutions, including Union Pro for Flux, which I've covered in the Flux ControlNet video. I'm very keen on the new Redux model, but it doesn't seem to work the way I have hoped. Anyhow, that's currently the best candidate for a video about the Flux tools.

What do you mean?

@@NextTechandAI I mean, I have spent over 2 days trying to get it to work and it will not. I have gone through various YT creator workflows and forget it. Ironically, I actually had XL almost do it, while the flux one next to it (from a creator) could not. Dev. I even tried your workflow to no avail.

@@generalawareness101 I still don't know what exactly didn't work for you, but in general you have to give the text enough space. Similar to finger inpainting, the new area to be inpainted needs to be large enough to actually accommodate 5 fingers.

@@NextTechandAI I gave it 1/4, 1/2, 3/4 of the images. I tried everything.

Sure, it just depends on the model and resolution you want to use. For low VRAM you can check my videos about GGUF for Flux and SD3.5. Both works best on AMD with Linux ROCm or Windows Zluda, see my related videos.

Thank you very much for the motivating feedback. I'm glad you found the video useful.

Hallo Kostya! Although I haven't used it yet, I think Amuse is a reasonable option to try out image generation. However, I think AMD should have put the resources into ROCm for Windows; ComfyUI, Forge and A1111 are proven open source tools that offer significantly more options. In my opinion, we don’t need another proprietary tool that is also based on ONNX.

Please post the contents of your directml bat. I assume you are calling the wrong bat or it is missing something.

@@NextTechandAI I reached the interface without any problems by doing the setup again. I checked the unet, clip and ae files and put them in their places. When I added the prompt to the queue and ran it, now I get the error "[F1119 23:44:35.000000000 dml_util.cc:118] Invalid or unsupported data type Float8_e4m3fn."

@@ancientlord5697 As I said in the video, Flux is currently not supported by directML.

​@@NextTechandAIIs there something i did wrong? Doesn't it work with directml in video?

@@ancientlord5697 No, I said in the video it doesn't work with Flux. I used Zluda after generating the example with SD15. You can use SD15, SD35 and SDXL.

Restarting PC helped for me because the Zluda dll didn't get recognized instantly in Environmentvariable

I'm glad you liked the video, thanks for the feedback.

Thank you for your feedback, I'm glad the video was useful for you 😀

I'm glad the video is useful. Which workflow do you mean?

Thank you very much, I‘m glad the tutorial is useful!

Thanks a lot. First inpaint the left character, in a second step inpaint the right one.

What do you mean? There is one VAE for Flux, but some checkpoints have it included directly.

@NextTechandAI dev has a special vae that can be downloaded on huggingface, maybe that is why the images turned out that poorly

@@as-ng5ln No, there is one VAE for Flux. This has absolutely nothing to do with the fact that Schnell follows prompts better than DEV. Try it yourself and generate the same image with both VAE files. By the way, you can try this with SD3.5 Large and Turbo, too.

@@NextTechandAI I'm telling you... I have the two files "ae.safetensors" and "flux1DevVAE_safetensors.safetensors". ae comes from schnell, while the other one is from the dev directory

@as-ng5ln Yes, and they have same effect on Flux image generation. As I said, try youself.

I followed my own video tutorial two days ago in order to install Ubuntu 24.04.01 on a USB drive. There haven't been any problems and it's working as expected. Are you choosing the installation target for the boot manager as described in the video?

Hi, which of my workflows and which model files do you use?

@NextTechandAI i use models SD. 3.5 Medium fp16 GGUF and i create manually same workflow on your this video but still getting error. i talk ChatGPT and update torch, transformers, diffusers and also setting torch float16 in my python not working aswell

@@PenkWRK Could you please use the suggested models and my workflows, you can download them from my patreon for free. In case this doesn't help, try the original SD3.5 Medium model without GGUF.

I'm glad you liked the vid.

VRAM refers to the GPU's memory, not the file size😉 Some have already gotten FLUX with GGUF to work with 4-6GB VRAM, I expect the same for SD3.5. With 8 GB or less, GGUF definitely makes sense, I also use it with 16 GB.

its a good amount but not for ai generation, i keep crashing the card when running SDXL models on a RX6600 lol 12 is defently on the edge of low vRAM for this type of stuff

These are checkpoints that I have downloaded over time and others that I have generated with Dreambooth.

By now WSL2 should be possible with an RX 7x00, my RX 6800 is still not supported.

An RX 6800.

I can understand that very well, I am very disappointed with AMD's support of AI, too.

It's the one mentioned in the description with "AMD ROCm on Windows Status Details and GIT & MiniConda-Installation".

@@NextTechandAI Thanks!!

For very low VRAM I've suggested in the video the FP8 T5, which is below 5GB. g and i together are about 1,5GB. You can even use the GGUF T5 encoders linked at the bottom of City96s GitHub with down to 2GB, but they have a bigger impact on quality than the quantized UNET models. Hence I'd try FP8 T5 first. Use runtime parameters --use-split-cross-attention and --lowvram or even --novram.

@NextTechandAI thanks I will have to try. Where do you use the parameters you mentioned? Do I use them when launching comfyui?

@@93simongh Yes, in the batch file directly after 'main.py'.

Thanks for your feedback. I'm happy that the plot community has another member.

I'm sure it's not too difficult to recognize my accent😉

If you can manage the installation with Zluda, then yes, extremely well.

@@NextTechandAI Everything works fine ?

@@louisbeauger There are a few restrictions, e.g. components using the bits-and-bytes extension do not work, like NF4, but I like GGUF much more anyhow.

@@NextTechandAI Ok Thank's !

From what I have seen there shall be a release soon. Forge with Medium 3.5 seems to be broken, probably they want to fix this first.

From what I have seen there shall be a release soon. Forge with Medium 3.5 seems to be broken, probably they want to fix this first.

Regarding quality or speed? Are you using both with GGUF?

Thank you!

Have you updated both the GGUF extension as well as your Comfy? Which GPU are you using?

@NextTechandAI I have updated everything and have a 12 GB RTX3060

@wolfgangterner7277 There is an issue in the GGUF github, which suggests several solutions: github.com/city96/ComfyUI-GGUF/issues/7 I think downgrading numpy as suggested in the bottom of this issue is the easiest solution.

Thank you very much!

Thank you. If you are using the models shown in my Flux GGUF video, I'd suggest Q8_0, Q5_K_S or Q4_K_S - the biggest one that fits.

Thanks a lot for your feedback.

Thanks a lot for your feedback. Interesting, I didn't know these two noes. Looks like by using them we can get something similar to 'masked only' in A1111. I don't think you need ControlNet for this. Not sure regarding upscaling, but usually it's a good idea to do at least a 2x after inpainting to blur the contours.

Thanks a lot! As mentioned in the video you can find the workflows on my patreon (for free). The link is in the description 😉

I published the first video of this channel on a German channel at the same time, it has around 600 views. The English version has almost 30k views. I'm afraid there is only a very small target group for such videos in German.

Thanks! Are you using a different version? I cannot remember a deepspeed option/attribute, can you possibly deactivate it?

​@@NextTechandAI I commented out these two lines referenced in the AttributeError and it seems to be working now: #if transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): #import deepspeed Location: stable-diffusion-webui\venv\Lib\site-packages\diffusers From what I can tell DeepSpeed is supposed to help accelerate the training. I may try installing it later but I'll work on getting my training tuned in first. My speed is adequate for now. Dreambooth extension version is 1b3257b4 (2024-08-04). Automatic 1111 v1.10.1 Python v 3.10.11 Torch 2.1.2 + cu121 xformers 0.0.23.post1 gradio 3.41.2

@vadar007 Thank you for the feedback, I'm glad it's working for you now.

how did you solve this?

@@AboodHani-t9r Sort the comments by Newest First and you'll see my reply in this thread that tells you how to fix it.

I'm happy you enjoyed the video😀

Thank you. In my tests there was always a slight difference, the ClipTextEncodeFlux seems to be better suited for T5. See my video about Flux prompting (ua-cam.com/video/OSGavfgb5IA/v-deo.html) regarding the two input fields. Frankly speaking I haven't seen ConditioningZeroOut quite often and it shouldn't have much influence, but from my point of view it looks more correct as Flux does not use negative prompts.