thanks for watching. MiniMax is working on image to video as we speak and will drop very soon. They are also upping the overall vid length from 6sec as well :D Fun Times
There is also another very big advantage of Juggernaut XI over flux. Speed. And yes, you can get faster flux, but with some big quality sacrifices. Flux is the iPhone and Juggernaut is the Android of the SD world. But they both have their plance in our hearts.
Appreciate the tutorial. I would be nice if you went at half the speed and made sure to clearly explain what you are doing at each step. You're rushing and it makes it hard to follow. You barely mention a step, flash a screen up, and then you move on.
Yes, normal Flux Dev supports it (not NF4) , I made a video on how to do it here - ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-MdMCGD1AZfE.html. Works quite well.
github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z Simply download, extract with 7-Zip and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints If you have trouble extracting it, right click the file -> properties -> unblock
Are you disrespecting Juggernaut by beeing ironic with your comparisons? Flux is the new standard. Nothing can compare yet. But I will always have fond memories of my days with juggernaut. (and still use it a lot because wel.. Ip-adapter and controlnet )
I wish , I still think Flux is much better overall as my comparison very limited. Hands text feet WAY better with Flux but SDXL still very useful ,(Apache 2 license) keen to test v12 (XII) and their Flux trained model when it comes out.
Excellent video and comparison. I think that at the moment Flux is very far from ideal. Yes, the anatomy is a little better, but everything else is currently at a weak level and requires a lot of improvements and training. Perhaps over time this model will reach the level of quality xl. And I'm not saying that the model is bad, no, it is not bad, it is different and this is the base on which you will have to work hard. And yes, I myself am trying to slowly work on improving this model, but it is very difficult.. Thanks for the video
Naa... best Model since long time: Dreamshaper Turbo SDXL. Best Model for Photorealism, only 6-9 Steps, superfast, 10 seconds on 8 GB VRAM in ComfyUI for 1024x1024. Upscaling very fast via Upscayle (Open-Source-Software). Flux sucks. Dreamshaper Turbo ist King.
You haven't used many SDXL models, have you buddy? Dreamshaper has to be one of the worst models available on SDXL. The creator has said multiple times that they aren't interested in SDXL as a model platform and in my experience anything that they have released for it is just half-assed. No offense, his SD1.5 model is one of the best, hands down. But the only thing his half-baked XL models have done is taken away potential downloads from creators who actually had the passion and drive to work with SDXL long-term.
I have my personal preference issues with older Juggernaut models, but no questions between JuggernautXL and DreamShaperXL which of the 2 projects was most heavily invested into. I have nothing but respect for the work that's been done on Juggernaut as the range of content it's been trained with effectively is beyond that of most models and it's prompt adherence is very decent. There are definitely better models, at least compared to older juggernaut versions I have tested. But most of them still fail to reach the same level of variation that Jugg has, only beating it in output quality or consistency, rather than knowledge base.
Forgot to mention I used flux dev for comparison. ALSO looks like Jug v12 (XII) will drop very soon too and rundiffusion (team behind Juggernaut models) look to have engaged black forrest for training with Flux model 😮 Happy days ahead 🎉
Are you using the same prompts for both? If so Juggernaut will come out better as you need to prompt differently for Flux so using it with SDXL style prompting is putting it at a disadvantage.
true. You can't test different models with the same prompting. ALso, he didnt mentionned which version of flux. Is that dev full or a water down version
thanks again for providing detail most of all the needed vram! i will have to check this out on weekend. I am still stuck (or in love) with the animate diff workflows...
got error with "Load and Resize" node issue with background_color it didnt like "image" so swapped it for basic image Load and that worked. (my bad, just needed to remove the word "image" from the original node setting for background_color)
So I'm trying to make a flux lora of my favorite character, but I want it to be very accurate from multiple perspectives. Would I just put different angles of the character in the zip file and name them correspondingly to what you said?
Great tutorial, it helped a lot. Could you please explain the 2 text boxes in the CLIPTextEncodeFlux. Do I write the same in both? Can I leave 1 empty? Is one the negative prompt? Why are there 2? And do you have a nice stress free workflow for hi res fix?
Thanks for the video, the gift, and the workflow. But running the Lora locally is painfully slow, like 1 hour and 15 minutes using your workflow on 8GB Vram.
My pleasure , you can try run flux dev NF4 for lower vram cards should be faster. Also could also try 512x512 resolution. Also adding --low vram tag and --fast tag to .bat might help
FLUX IPADAPTER & CONTROL NETS came with implicit restrictions as censorship? for example, could make a nsfw model became nsw? thanks Also appreciated you shared this video
I created a workflow leveraging Ollama to generate a prompt and then share that prompt with the Process Node and a SDXL image generation which sends the image to the "controlnet_image" input. Setting the Controlnet_weight = 1.0 it is pretty close to the original image with amazing details and upscaling. The initial image is 1344x768 and is upscaled to 3840x2304 which allows me to make the images a wallpaper. Works very well. Falling in love with this.
Isn't there an IA that automatically changes sharpness/blur, contrast, bright, light, shadow, black, white and saturation in the Mona Lisa picture until the control net increases to a satisfactory quantity, before proceeding to generate the final image?