Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
The workflow is attached as RAR - it is inside the image.
You don't have permission to view the spoiler content.
Log in or register now.
Last edited:
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
So far it seems Qwen understands woman 1 and woman 2 and responds to them appropriately:
Here prompt reads: Woman 1 sits on a stool, woman 2 stands with hands behind her back.
But! It didnt work at all when I swapped the numbers: Woman 2 sits on a stool, woman 1 stands with hands behind her back.
Girl 1 still sits on the stool, girl 2 is behind. Hmm, not as straight forward as I imagined.
Here prompt reads: Woman 1 sits on a stool, woman 2 stands with hands behind her back.
But! It didnt work at all when I swapped the numbers: Woman 2 sits on a stool, woman 1 stands with hands behind her back.
Girl 1 still sits on the stool, girl 2 is behind. Hmm, not as straight forward as I imagined.
Last edited:
JhonLui
Well-Known Member
- Jan 13, 2020
- 1,189
- 1,178
- 284
Try to work it the other way around..
Instead of swapping position in the prompt, swap the description (and lora if any) or just the connection if I read the diagram right..
Sharinel
Active Member
- Dec 23, 2018
- 614
- 2,444
- 448
Hmm, maybe try woman in red cloak sits on stool, women in black bikini stands behind? Something like that?
- Jul 4, 2017
- 3,411
- 6,565
- 707
I don't read Chinese but from my understanding of Kanji from Japanese it's just very basic:
"the woman in image 2 is carrying on shoulder the bag in image 1"
"the vehicle in image 1 is stood next to by the woman in image 2" (grammar is obviously borked when doing original sentence order)
The third one about the wedding shots is way too complex for me.
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
What had my jaw on the floor, an "aha", or "we have arrived" moment is converting from cowboy shot to full-body shot.
This new Qwen is a major upgrade over the original one. Naturally, there are these issues with size and color consistency, these are not lost on me, be alas, the trajectory of the development will take us to new levels of depravity.
Original | Get fullshot view | Legs spread | Put clothes on |
|
|
|
|
This new Qwen is a major upgrade over the original one. Naturally, there are these issues with size and color consistency, these are not lost on me, be alas, the trajectory of the development will take us to new levels of depravity.
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
Very few things in life are as fun as Qwen 2509.
And we literally have a new era in AI - a whole new toolkit that today allows something that wasn't available yesterday.
So far only one major issue -- Qwen is much slower than say Illustrious, and upscaling it on my machine using model-upscaler takes around an hour.
And we literally have a new era in AI - a whole new toolkit that today allows something that wasn't available yesterday.
So far only one major issue -- Qwen is much slower than say Illustrious, and upscaling it on my machine using model-upscaler takes around an hour.
- Jul 4, 2017
- 3,411
- 6,565
- 707
What is "model-upscaler" and what does it do? I don't understand how a model can be upscaled.
Or are you saying that your comfy pipeline has a step after qwen image generation that performs a (e.g.) 1k->4k upscale, and the whole operation takes around an hour?
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
Yea, it is this guy - "Ultimate SD Upscale" node - pardon my pedestrian calling it "model-upscaler". IDK why I do that, probably to contrast a simplier upscaler that doesnt use a model to upscale. It is probably a tile controlnet under the hood.
It gets me to say 2000x3000 in Illustrious in about two minutes and looks great:
Last edited:
JhonLui
Well-Known Member
- Jan 13, 2020
- 1,189
- 1,178
- 284
can I ask you what machine you have? just to have an idea on what it would take on mine..
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
JhonLui
Well-Known Member
- Jan 13, 2020
- 1,189
- 1,178
- 284
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
Here's 2025 in review*.
*2025 in review so far.
If you are still on a fence about Qwen 2509, let me show you lil something.
1. Use Illustrious to generate a reference image. There's a gazillion of references with workflows on Civitail - anything you can possibly want.
2. Use Qwen to transform the reference into a "standard look" - ask for "full body shot" and the attire you want your talent to sport. 90% of work I do in Qwen uses 1.1 Lightning LORA, including all images in this post.
I still debate (will be doing more tests on this), whether you want multiple views from all sides already generated so you can pick and choose the one that fits the scene the most:
I tried using character sheets, etc, but specialist images tend to give the very best result. Character sheets give the best results - you want "front view, back view" only, don't include side view with the sheet. Also, you want your talent already dressed for the scene, i.e.:
3 & 4. Deploy your talent into your environment. Guess which environment image was used for the "girl & lockers" set? The answer: D.
I will be doing more tests to check if Qwen prefers D-like images for composition over A, B and C.
BTW, all these school halls / lockers are Qwen-Edit generated. I.e. environment A came free when I asked to put two girls (from another set) into a school hall, so Qwen gave me this which I cleaned up using Qwen:
B, C and D were acquired in the same way.
Now, an important point about the final image: aesthetics of portrait vs landscape orientation are very different. It seems 1024x1280 (WxH) are the sweet-spot, then 1280x1024 (WхH) even tho the official release page suggests 1024x1024.
Issues, etc.
1. There is still a massive issue that I don't know how to overcome - relative sizes. If you want two charas in your image, then getting their sizes right relative to each other is not gonna work. If anyone finds a consistent solution, I hope that hero would share it in due time.
2. Nowhere near Illustrious in terms of things like "cameltoe", "see through", etc, etc - definitely missing those. But! Oh boy did we improve over the original Qwen Edit -- 2509 is like 10000% better about all things smut.
3. Plastic look. I imagine this is the issue of my workflow where the same talent gets Qwen'ed multiple times inevitably loosing their natural look. Something for later.
2026 might turnout to be lit, off the hook and off the wall.
*2025 in review so far.
If you are still on a fence about Qwen 2509, let me show you lil something.
Generate a Reference | Clean Up & Dress Up | Deploy 1 | Deploy 2 |
|
|
|
|
| Full body shot. Front view. She is naked, wears white heels. Hands lowered. Black background. | Put woman in school hall. Rear view of woman who stands against lockers. Looks at viewer. Blend the lighting so it looks natural. | Put woman in school hall. Rear view of woman reaching into an open locker. Looks at viewer. Blend the lighting so it looks natural. | |
--- | --- | --- | --- |
|
| [TBA] | [TBA] |
| Full body shot. Front view. Wears blue converse. Black background. | [TBA] | [TBA] |
1. Use Illustrious to generate a reference image. There's a gazillion of references with workflows on Civitail - anything you can possibly want.
2. Use Qwen to transform the reference into a "standard look" - ask for "full body shot" and the attire you want your talent to sport. 90% of work I do in Qwen uses 1.1 Lightning LORA, including all images in this post.
I still debate (will be doing more tests on this), whether you want multiple views from all sides already generated so you can pick and choose the one that fits the scene the most:
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
I will be doing more tests to check if Qwen prefers D-like images for composition over A, B and C.
| A | B | C | D |
|
|
|
|
BTW, all these school halls / lockers are Qwen-Edit generated. I.e. environment A came free when I asked to put two girls (from another set) into a school hall, so Qwen gave me this which I cleaned up using Qwen:
You don't have permission to view the spoiler content.
Log in or register now.
Now, an important point about the final image: aesthetics of portrait vs landscape orientation are very different. It seems 1024x1280 (WxH) are the sweet-spot, then 1280x1024 (WхH) even tho the official release page suggests 1024x1024.
Issues, etc.
1. There is still a massive issue that I don't know how to overcome - relative sizes. If you want two charas in your image, then getting their sizes right relative to each other is not gonna work. If anyone finds a consistent solution, I hope that hero would share it in due time.
2. Nowhere near Illustrious in terms of things like "cameltoe", "see through", etc, etc - definitely missing those. But! Oh boy did we improve over the original Qwen Edit -- 2509 is like 10000% better about all things smut.
3. Plastic look. I imagine this is the issue of my workflow where the same talent gets Qwen'ed multiple times inevitably loosing their natural look. Something for later.
2026 might turnout to be lit, off the hook and off the wall.
Last edited:
Sharinel
Active Member
- Dec 23, 2018
- 614
- 2,444
- 448
2. Does it do anything like penetrative sex?. Also do you have a workflow that you can attach? I'll have a look in between my wan 2.2 splurge
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
I haven't tried anything about penetration yet, I do not know.2. Does it do anything like penetrative sex?. Also do you have a workflow that you can attach? I'll have a look in between my wan 2.2 splurge
Here's a QE 2509 workflow: https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-18185892
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
Maybe you already saw this, but here in about 30 sec a dude explains how Qwen supplements Wan 2.2.2. Does it do anything like penetrative sex?. Also do you have a workflow that you can attach? I'll have a look in between my wan 2.2 splurge
Basically Qwen gets you the start and end frames, and Was does the heavy lifting.
If you haven't seen, here, starts at 9:29 and goes for abour 30 sec:
You must be registered to see the links
I was playing aperture. an h-game that uses AI image.
I was impressed with this image.
Image models are trained with single character portrait images and are very bad at background and wide shots images.
How do you think this was generated?
I tried to generate a similar image. I was able to get the poses of the characters right but the background composition was very bad with the view was always close to the characters instead of a wide shot and the background street with low-height buildings very bad and the haze in the sky also very bad.
This is the image from the game:
This is the image I generated:
I was impressed with this image.
Image models are trained with single character portrait images and are very bad at background and wide shots images.
How do you think this was generated?
I tried to generate a similar image. I was able to get the poses of the characters right but the background composition was very bad with the view was always close to the characters instead of a wide shot and the background street with low-height buildings very bad and the haze in the sky also very bad.
This is the image from the game:
This is the image I generated:
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
Idk how it was done exactly, but going forward wide-angle composition will be done using Qwen 2509:
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,709
- 4,192
- 448
Wan 2.2's start-frame - end-frame is a "whoa". Literally sitting here realizing I can make an entire hour long smut video in my (ahtually not my) basement.
Fellow "artists", if you haven't tried W22 yet, then pull the finger out already and go for the frame-to-frame workflow, tinker with it. Just avoid the Lora 4-step lora - while lora lets you have videos fast they are garbage compared to what raw W22 can give you. There are probably use cases for the lora, but avoid it until we figure out what they are. Trust me the 30 min wait with the raw W22 is absolutely worth the ~5sec clips it gives you.
Start with 640x640 resolution, crop your reference frames to 640x640 and go with default 81 length and 16 fps. After even a few hours of tinkering you'll realize what a powerful tool this is: it is up there with bunker busters and compound interest.
Naturally, Wan 2.2 and Qwen 2509 are just made for each other - if you have solid experience in both you are literally getting promoted to a porn movie director in a quick minute.
Fellow "artists", if you haven't tried W22 yet, then pull the finger out already and go for the frame-to-frame workflow, tinker with it. Just avoid the Lora 4-step lora - while lora lets you have videos fast they are garbage compared to what raw W22 can give you. There are probably use cases for the lora, but avoid it until we figure out what they are. Trust me the 30 min wait with the raw W22 is absolutely worth the ~5sec clips it gives you.
Start with 640x640 resolution, crop your reference frames to 640x640 and go with default 81 length and 16 fps. After even a few hours of tinkering you'll realize what a powerful tool this is: it is up there with bunker busters and compound interest.
Naturally, Wan 2.2 and Qwen 2509 are just made for each other - if you have solid experience in both you are literally getting promoted to a porn movie director in a quick minute.
Last edited: