imo, the best approach would be to generate a bunch of random images using your favorite SD1.5 LoRA and then train a Flux 1.D LoRA with these images. This could be done easily now. - You're essentially enhancing a SDXL image with FLUX. So you're mainly losing out on the main advantage of FLUX: which is prompt adherence. - If you're after high-quality / detailed images, consider using SDXL, with HiresFIX and/or creative upscaling, and/or multistage generations. These methods can outperform a "Flux enhancing" when it comes to refining your SDXL images. Anyway, great channel. Thanks for sharing. 😊
I integrated Qwen2 VL and it works much better than Florence. Great video by the way. I think a lot of people are looking for these kinds of tutorials. Step by step, east to understand and most importantly - WHY. Great job.
This is the similar concept of workflow I use currently. Use Flux for refining the image before upscale in the final step. SD1.5 or SDXL --> Flux --> Upscale as final step.😆
@@CodeCraftersCorner Hello friend, could you make a workflow with flux. With inpaint adding flux Loras, I've been looking for a long time but I can't find it, I want to do inpaint with a Lora that trains for flux
Thank you for posting this, and right at the exact time I was interested in doing something similar. This is awesome! I'll certainly be following your vids. However, I find it difficult to follow your connections (noodles) when they are straight lines and hidden underneath node boxes or when they twist behind the node they come from leaving very little to visually follow. I realize I could just copy your workflow but I'm trying to learn the logical steps and reasons for creating and using each type of node. It is really your commentary that is the most helpful My biggest concern is that this is pointless. As I understand it, LoRAs are meant to enhance or add to what Checkpoints fail to do. By using an SD or SDXL LoRA before the FLUX model, anything the LoRA generates but is not in the training data of the FLUX model, will be washed away. If I use an SD LoRA of a specific artist's style the final image after FLUX will be nothing like that artist's style.
Thanks for sharing your efforts, I'll give it a try. I am stuck with my poor old GTX 1070, which can manage Flux fp8 but doesn't have the vram to handle a lora in a Flux workflow, so this may be the solution.
I'm having trouble loading the Florence-2-base model (node Florense2ModelLoader). Which items should I download correctly so that they work with the corresponding node? In which Comfyui folder should the items be dumped?
Hello. You can use the "Download and Load Florence 2 model" node to automatically download the model. You can check the video at timestamps: 04:46 for assistance.
@@CodeCraftersCorner Thank you very much for responding. I'm from Brazil and I speak Portuguese, tutorials in other languages can only be done with Google's automatic translation. But since ComfyUI is by nature confusing, even if you speak slowly in your videos, something always gets lost. Hugs!
Hello, the load image is a default node. It should be available at all time. Can you try to manually add it? You can also try to update ComfyUI through the manager and restarting ComfyUI.
hey! for some reason the florence 2 model is the 1 and only node that wont install for me. im not sure why. I hit install missing nodes, updated comfyui, hit try update on the node multiple times, restarted comfyui multiple times, nothing is working. every time I try to load the workflow just that 1 node is missing. :( (very top left node)
Hello, can you try to download the model with this node first: "Download and Load Florence 2 model". You can check the video at timestamps: 04:46 for assistance.
I feel compelled to ask why not use other models to achieve the result you want? You are literally trying to duplicate an image you have. And using 1.5 to gen it… If you want more details use a detailer and upscaler
That's right! If you're getting good results, that's what matters. As I mentioned, I'm integrating SD LoRAs with the Flux workflow since I no longer use the SD workflow. In the end, it comes down to preference and what works best for you. After all, detailers and upscalers can be used with Flux as well.
Flux Loras are best especially for person. It would be great to be able to use Flux Loras in SD1.5 or SDXL. I'm going to try your method backwards and see if it works. Edit: It didn't work, I got similar picture but with a different person.
imo, the best approach would be to generate a bunch of random images using your favorite SD1.5 LoRA and then train a Flux 1.D LoRA with these images. This could be done easily now.
- You're essentially enhancing a SDXL image with FLUX. So you're mainly losing out on the main advantage of FLUX: which is prompt adherence.
- If you're after high-quality / detailed images, consider using SDXL, with HiresFIX and/or creative upscaling, and/or multistage generations. These methods can outperform a "Flux enhancing" when it comes to refining your SDXL images.
Anyway, great channel. Thanks for sharing. 😊
This is indeed a better long term solution. Thanks for sharing!
I integrated Qwen2 VL and it works much better than Florence. Great video by the way. I think a lot of people are looking for these kinds of tutorials. Step by step, east to understand and most importantly - WHY. Great job.
Thanks for sharing! Glad you liked them!
This is the similar concept of workflow I use currently. Use Flux for refining the image before upscale in the final step. SD1.5 or SDXL --> Flux --> Upscale as final step.😆
Thanks for sharing!
This is so far the best approach. Thanks!😁
Thank you!
Nice tutorial
Thank you!
Very nice tutorial. You should do more comfy tutorials. Thanks for doing this one. :)
Thank you! Will do!
Amazing!!! Thank you bro!
Glad you like it!
Respect for consistent detail explanation as always, Thank You
Much appreciated
Thank you for sharing your workflows!
My pleasure!
@@CodeCraftersCorner Hello friend, could you make a workflow with flux. With inpaint adding flux Loras, I've been looking for a long time but I can't find it, I want to do inpaint with a Lora that trains for flux
thank you for the great explanation!! very clear!
Glad you enjoyed it!
Thank you for the amazingly detailed video!
Glad you enjoyed it!
Thank you for posting this, and right at the exact time I was interested in doing something similar. This is awesome! I'll certainly be following your vids.
However, I find it difficult to follow your connections (noodles) when they are straight lines and hidden underneath node boxes or when they twist behind the node they come from leaving very little to visually follow. I realize I could just copy your workflow but I'm trying to learn the logical steps and reasons for creating and using each type of node.
It is really your commentary that is the most helpful
My biggest concern is that this is pointless. As I understand it, LoRAs are meant to enhance or add to what Checkpoints fail to do.
By using an SD or SDXL LoRA before the FLUX model, anything the LoRA generates but is not in the training data of the FLUX model, will be washed away.
If I use an SD LoRA of a specific artist's style the final image after FLUX will be nothing like that artist's style.
Thanks for the input.
Thanks for sharing your efforts, I'll give it a try.
I am stuck with my poor old GTX 1070, which can manage Flux fp8 but doesn't have the vram to handle a lora in a Flux workflow, so this may be the solution.
Glad to hear this could be a potential solution.
Great video!
😊
Glad you liked it!
I'm having trouble loading the Florence-2-base model (node Florense2ModelLoader). Which items should I download correctly so that they work with the corresponding node? In which Comfyui folder should the items be dumped?
Hello. You can use the "Download and Load Florence 2 model" node to automatically download the model. You can check the video at timestamps: 04:46 for assistance.
@@CodeCraftersCorner Thank you very much for responding. I'm from Brazil and I speak Portuguese, tutorials in other languages can only be done with Google's automatic translation. But since ComfyUI is by nature confusing, even if you speak slowly in your videos, something always gets lost. Hugs!
Very interesting !!
Glad you think so!
When I try to use your workflow, the "Load Image" option isn't appearing. Am I missing something? Thanks!
Hello, the load image is a default node. It should be available at all time.
Can you try to manually add it? You can also try to update ComfyUI through the manager and restarting ComfyUI.
Hi Sharvin - Do you know if there is any way to use QRCode monster controlnet with flux to create surreal images?
Hi Sebastian, I will check and let you.
hey! for some reason the florence 2 model is the 1 and only node that wont install for me. im not sure why. I hit install missing nodes, updated comfyui, hit try update on the node multiple times, restarted comfyui multiple times, nothing is working. every time I try to load the workflow just that 1 node is missing. :( (very top left node)
Hello, can you try to download the model with this node first: "Download and Load Florence 2 model". You can check the video at timestamps: 04:46 for assistance.
Why are the lines connecting the nodes in your workflow so straight and neat? Normally the lines between nodes look like spaghetti strings.
you can change it in the settings, it's not hard to find
Yes, you can change it from the settings (gear icon). Search for render lines or line and you should get it. Change the option to straight.
Isnt this just an image to image workflow? Or have i missed something?
Yes, it uses image to image.
what about using canny from the original image
Thanks for the idea. I will try with ControlNets.
I feel compelled to ask why not use other models to achieve the result you want? You are literally trying to duplicate an image you have. And using 1.5 to gen it… If you want more details use a detailer and upscaler
That's right! If you're getting good results, that's what matters. As I mentioned, I'm integrating SD LoRAs with the Flux workflow since I no longer use the SD workflow. In the end, it comes down to preference and what works best for you. After all, detailers and upscalers can be used with Flux as well.
Flux Loras are best especially for person. It would be great to be able to use Flux Loras in SD1.5 or SDXL. I'm going to try your method backwards and see if it works.
Edit: It didn't work, I got similar picture but with a different person.
Thanks for sharing your results. Also agree on Flux loras for person.
watching at 1.5x
😊
Glad you're enjoying the video at 1.5x! I do the same-actually made my own Chrome extension for it too. 😀