Phil2Sat

/

Qwen-Image-Edit-Rapid-AIO-GGUF

like 227

Based on my observations so far, v5 is currently the best. I noticed that v5.2 significantly alters faces when editing photos. It exaggerates body contours in an unnatural way. It's disproportionate. v5.3 is currently in the download phase.

Based on my observations so far, v5 is currently the best. I noticed that v5.2 significantly alters faces when editing photos. It exaggerates body contours in an unnatural way. It's disproportionate. v5.3 is currently in the download phase.

https://civitai.com/models/1939453/qwenedit-consistence-lora this can slightly fix this, but what i observed:

• loras apply different on FP8 or Q quantized models

So in my testing phase i tried to recreate all the loras from 5.2 and apply them to the stock 2509 gguf. Result should be the same with same weights. But its not.

• Get weird embossed images out of it, so i thout about why, its the dynamic range i guess, its lower on FP8, cause my GGUF uses F16/Q_8 blocks and they are very similar think about Q8 is nerly f16 and FP8 is half of that

• So if you apply a lora which is trained on FP8 to F16/Q_8 with weight 1 it oversteers heavy.

• lets take the qwenedit-consistence-lora recommended weight is 0.48 but it gives gridlines, now we think about the half dynamic range, so applying it with something between quarter to half weight gives good results. so somewhere between 0.12 to 0.24 should produce the results wanted.

• next thing from v5 to v5.3 two extra loras were added. every new lora wants to change the result the way it was trained, qwen base is chinese, so alot of chinese training data, https://civitai.com/models/2058077/qwen-imagensfwadv1 is western so alot of western training data.
  -Keeping faces consistent is something i wasnt able to reach at all, no matter if i tried sdxl, flux or qwen. I guess the only method to keep faces really consistent is to train a character lora and flood the model with thousands of training images of that exact face.

here is a actual test, thats what i got with 5.3 and consistence 0.24 4-step BTW one of my first images with 5.3
The face is squished, it looks always younger, this is a generated input taken somewhere but if i take a portait of my wife, thats never ever my wife, no matter what model i take, trying since a month or so.

Based on my observations so far, v5 is currently the best. I noticed that v5.2 significantly alters faces when editing photos. It exaggerates body contours in an unnatural way. It's disproportionate. v5.3 is currently in the download phase.

https://civitai.com/models/1939453/qwenedit-consistence-lora this can slightly fix this, but what i observed:

• loras apply different on FP8 or Q quantized models

So in my testing phase i tried to recreate all the loras from 5.2 and apply them to the stock 2509 gguf. Result should be the same with same weights. But its not.

• Get weird embossed images out of it, so i thout about why, its the dynamic range i guess, its lower on FP8, cause my GGUF uses F16/Q_8 blocks and they are very similar think about Q8 is nerly f16 and FP8 is half of that

• So if you apply a lora which is trained on FP8 to F16/Q_8 with weight 1 it oversteers heavy.

• lets take the qwenedit-consistence-lora recommended weight is 0.48 but it gives gridlines, now we think about the half dynamic range, so applying it with something between quarter to half weight gives good results. so somewhere between 0.12 to 0.24 should produce the results wanted.

• next thing from v5 to v5.3 two extra loras were added. every new lora wants to change the result the way it was trained, qwen base is chinese, so alot of chinese training data, https://civitai.com/models/2058077/qwen-imagensfwadv1 is western so alot of western training data.
  -Keeping faces consistent is something i wasnt able to reach at all, no matter if i tried sdxl, flux or qwen. I guess the only method to keep faces really consistent is to train a character lora and flood the model with thousands of training images of that exact face.

here is a actual test, thats what i got with 5.3 and consistence 0.24 4-step BTW one of my first images with 5.3

I downloaded v5.3 and started using it. I edited higher quality and more proportionate images than with v5.2. The disproportionate and constant face-changing problem in v5.2 seems to be solved for now.

Using : V5.2 Q6 VS V5.3 Q6

I compared NSFW v5 to v5.3 (both Q4_K_M) and found v5.3 to be much worse, image contrast change, some have gridlines, and also hand position changes. v5 had similar image tone to original, only some images had problems with hand placement issues.

I compared NSFW v5 to v5.3 (both Q4_K_M) and found v5.3 to be much worse, image contrast change, some have gridlines, and also hand position changes. v5 had similar image tone to original, only some images had problems with hand placement issues.

what settings? maybe try euler_a beta/sgm_uniform
large changes sometimes need one or two steps more.
did you add additional loras? if yes try half weight.

tink about that every additional lora needs some percentage of a step for changes, with alot of loras +3 in v5.3 it could need more steps to bring in every detail all the lowas want to add. at least my brain translate it that way

My settings: 4 steps, No Loras, eular_a, sgm_uniform, Qwen2.5-VL-7B-Instruct-abliterated, pig qwen vae. I've tried 4steps. 8steps, changed schedulers and samplers, swapped to Qwen_image_vae, the image tone definitely changes into more red tone.

ah the tone, try cfg norm node with 0.92.

but did the gridlines disapear at higher steps?

te q4_k_m
v5.3 q5_k_m
5-steps euler_a beta
edit + pose transfer (openpose included but not needed)

including workflow:

for a little bit more realism let the NSFW kick in add some bad words:

hey Phil, great workflow. i wonder what is this node and where can i find it ? thanks ;D

Its not needed simply remove, its for something if you want to pre transform a position of a person, maybe its sitting and you want a laying position, so its easier for qwen in a few steps to get in position.

but if you want, MixLab https://github.com/shadowcz007/comfyui-mixlab-nodes, i have such a bunch of nodes installed, i dont really remember what is what, have to look them up, hehe

Phil, im sorry again to bother. Skill issue from my behalf, I was wondering why, the fingers doesnt' get interpretated ? Sorry this isn't the place to ask, but im still learning, a lot to process in so little time.
But, to honor the main topic, I include my artwork (would not call this artwork, just a test!) using Q6_K

Try changing the openpose to dwpose node:

im also testing right now and dwpose gives better fingers.

just testing different things...