Phil2Sat

/

Qwen-Image-Edit-Rapid-AIO-GGUF

like 254

actually uploading a refresh so in first step v5.2, after that an optimized quant of v5.

wanted to cleanup folder structure so deleted everything

i guess i skip 5.1 completely

takes a while to upload 220gb with 50mbit

thanks love your stuff hope you do a flux AIO one day

I think skiping v5.1 is ok. Thank you for keeping this repo up to date :)

We're eagerly awaiting it! Thank you!

97/226GB at ~7mb/s, hate the slow upload and next the updated v5, another ~226gb.

dang just saw v5.3 released, that will be a long upload session today:
for what i saw:
added https://civitai.com/models/2058077/qwen-imagensfwadv1
and reeanabled cockqwen some minor weights change like from 0.8 to 0.75 for mncl, meta4 and qwen4play loras

dang just saw v5.3 released, that will be a long upload session today:
for what i saw:
added https://civitai.com/models/2058077/qwen-imagensfwadv1
and reeanabled cockqwen some minor weights change like from 0.8 to 0.75 for mncl, meta4 and qwen4play loras

Yeap 5.3 released. I think we'll see more high-quality models.

btw developed a new quant method, sorting exactly which blocks needed to stay in high precision, landed somewhere between my first Q8_0 attempt and the hq version i took offline which had some blocks in fp32.
since im testing with the worst possible quant q2_k i have to say i see no real difference in the original model compared to the q2_k.

neat thing future uploads are a little bit smaller with a little bit better quality and prompt adherance. at least my yesterday tests showed me.
left 4-step q2_k v5.2, right input, added https://civitai.com/models/1939453/qwenedit-consistence-lora with 0.48 additionally in my comfyui workflow on top of the AIO.

this is a cutout of a from thighs, 720x1456 image without any post process, i think for the nah i wouldnt use any q2_k it does well.

v52 up
v50 uploading
v53 gguffing

v52 up
v50 uploading
v53 gguffing

YAY FINNALLYY! Is there much difference between 5.2 and 5.3? I'm downloading 5.2 right now.

couldnt even test 5.3 as my machine is working... i guess until tomorrow

BTW 5.2 Q5_K_M Not working as well. The model isn't doing any of what I wrote. I think there's a problem.

NVM, It's working this is my fault lol.

all up..

v7.1 up, initial tests, bad at multi image and consistency, much worse than v5.3

Agree, too focused on NSFW content. Bad anatomy on heads (too big), not natural looking. Will try some more and maybe report back.
Noticed now: "V7: valiantcat/Qwen-Image-Edit-MeiTu and chestnutlzj/Edit-R1-Qwen-Image-Edit-2509 included as LORAs.".
That MeiTu version does not work good. Tried it earlier also.

v7.1 up, initial tests, bad at multi image and consistency, much worse than v5.3

I think it gets worse as more lora is added. For now, v5 is the best for me.