Edit Images with Flux RF-Edit, Regional Prompting & PAG in ComfyUI with 8+GB VRAM

RF Inversion (unsampling) works well, but RF Edit gives you another tool in the toolbox for making image edits. Flux Regional Prompting lets you prompt specific regions, with PAG/SEG adding extra details and a little prompt adherence to your generations. Best of all, you get all three from the Fluxtapoz GitHub repo!

Run for free on your own computer at home, and let your imagination run wild!

Links:

Want to support the channel?

== Beginners Guides! ==

== More Flux.1 ==

Contents
---------------
0:00 Fluxtapoz RF-Edit Workflow
6:38 Fluxtapoz InstantX Regional Prompting
12:07 Flux Perturbed Attention Guidance (PAG) and Smoothed Energy Guidance (SEG)

[juanjesusligero]

Oh, Nerdy Rodent! 🐭🎵
He really makes my day! ☀
Showing us AI, 🤖
in a really British way! ☕🎶

[mond]

These are powerful techniques. AI image generation is advancing so quickly! Thanks for keeping us up to date.

[Guus]

Thank you for the great showcase! Really helpfull <3

[Sandy-D-]

I'll understand that title if I watch it, right? Right?

[Paulo-utli]

Great video! About rfedit slowness, at least with 4090, using force clip cuda:0 and dual clip gguf with t5 q5 decreased the generation time by half with rfedit.

[wakegary]

Oh, Nerdy Rodent! 🐭🎵
He really makes my day! ☀
Showing us AI, 🤖
FOR BRITISH EYES ONLY!!!

[steve_jabz]

Does regional prompting work with LoRas?

[katonbunshin]

Why u removed video : Hunyuan Video in ComfyUI with 8GB+ VRAM ?

[bgtubber]

PAG still makes the generation times longer. So be aware, especially with Flux being already slow to run even on top-of-the-line hardware.
BTW, in my initial tests, Flux PAG has zero effect for me even at scale set to 4.0 in an img2img workflow with the Ultimate SD upscaler, Shakker's Union Pro tile controlnet, and the "PixelWave" Flux model.

[CharlesLijt]

Man, 12:44 how did you get such different result while enabling PAG attention node? My PAG does NOT change the image at all, very very little change in the image, hardly any little detail at all, I saw someone mentioned in the ComfyUI-Fluxtapoz issue page about this situation as well...

[Spinaster]

Hi, and thank you for the tutorial! It inspired me to create something similar, with the ability to generate multiple masks directly using the Mask Editor and Separate Mask.
After spending hours trying to understand why I couldn’t achieve results similar to the ones you showed, I realized that the Redux Advanced node probably doesn’t work as it should.
By replacing it with the Regional Style Model Apply and Clip Vision Encode, I get the desired results perfectly. The style is transferred much more consistently.
Additionally, I’m still investigating, but it seems that the Create Flux Regional Cond node also doesn’t work as it should. From the tests I’m running, it seems to interpret the black areas of the mask as those to “fill, ” which is the exact opposite of how a mask should be interpreted.
I’m not entirely certain about this because the original mask passes through the Grow Mask with Blur node, which could be the problem.
Let me know what you think!

[ChanhDucTuong]

I disabled Xformer, the workflow works but it only give me a big black image, any idea please? Before disabling Xformer the workflow still work but also only give me black image.

[Main]

Fluxtapoz non-deterministic fix is to move line 10 in rectified_sampler_nodes to before line 16 where y1 is made.

[Kevik]

I have replicate the RF edit workflow but I only experience minor changes but only injecting steps on 0 produce any significant changes, which is why I need to play around the unsampling denoise strength, which doesnt produce great results.

What could be the issue here?

[sheldonventer]

I would love if you could also cover outside of Flux, like the anime illustrious, There is an LLM for that uses Danbooru tags and natural language and mashes them together, man I would love if you could cover this, the LLM thing is for KBlueLeaf/ TIPO. I cant get it to work. x_x

[Zeclink]

Thanks for the great showcase!
It was really helpful.
Could you guide me on how to replicate the workflow shown in your video?

[MyMrJensen]

I'd love to understand why SEG and PAG seem to work for you ... All I get is increased run-time when using any of them, but with the exact same results