[Stable Diffusion] Prompt Sharing and Learning Thread

[JhonLui]

Wan 2.2's start-frame - end-frame is a "whoa".

Go take a look to MagCache if you haven't already. it promises to speed up 4x.
I'm starting to get a hang on it on Hunyuan and Framepack Studio and (depending on what you need) it goes as fast as 2sec/it on a 4060Ti 8GB (average for a good res realistic 5sec/it)

 

[Sharinel]

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.

View attachment 5294130

I am using the 4-step lora (sorry!) so this may be based on that, but I have found that fp of 16 makes everything now and again run in slow-mo, while 24fp makes them run about like Benny Hill being chased, 20fp seems to work quite well for me

 

[osanaiko]

Benny Hill

I see you, old man.

 

[Sepheyer]

I am using the 4-step lora (sorry!) so this may be based on that, but I have found that fp of 16 makes everything now and again run in slow-mo, while 24fp makes them run about like Benny Hill being chased, 20fp seems to work quite well for me

Indeed, I've noticed the same - lora removes the annoying dynamism the raw model has. Even when I ask the talent to stay still she still clowns around like Chihuahua on speed.

But here's an issue I have with lora - it fumbles simple tasks and has to do crossfade. May be this is me getting something wrong. Let me show you what I mean - both vids are zoom in / zoom out and yet lora can't handle it without a crossfade. In fact I see a lot of crossfades with lora. Yet the raw wan handles the transition very well, altho the bimbo acts larger than life.

LoRA	Regular
View attachment a_00026_.mp4	View attachment a_00022_.mp4

In the middle of this post I realized I should probably take a "still" video of her and just zoom in manually in a video editor, lmao.

 

[osanaiko]

Indeed, I've noticed the same - lora removes the annoying dynamism the raw model has. Even when I ask the talent to stay still she still clowns around like Chihuahua on speed.

LORA aside, this same sort of "annoying dynamism" aspect is something that is *annoying me* when using Flux.1-Kontext for img2img work.

Despite using "slightly" or "a little" or "subtle" modifiers, the img generator just *cannot* manage to do anything other than a 120% version of every expression or pose requested.

"Change the woman's expression to slightly upset. Keep the same facial features, hairstyle, colors, and clothing." -> girl looks like you just tore the throat out of her kitten with your teeth, while everything else in the image is exactly what I wanted.

I've resorted to going back to photoshopping eyes / mouth etc to be smaller or less extreme as a postwork step. It's still way better than my non-existent ability to draw human features, but it is incredibly annoying that we can get so close but still so far to the tools being a true magical assistant.

 

[Sepheyer]

Qwen 2509 keeps growing on me. It is THE composition tool: you make a waify with Illustrious, then you put her thru the scenes with Qwen.

I spent quite a good teal of time with Illustrious and SD to know that Qwen brings an entirely new capability for pose control right out of the box. Granted, Qwen is one massive control next under the hood, but still.

Also check out the perspective with regard to full body shots and how natural it looks - this is new (at least to me).

FML, you are literally self-promoted to chief editor / senior photographer at Playboy circa 1985.

Make sure to employ an 8-step lora and "Patch Sage Attention" to cut down ~30% on generation time.

Also, don't use Qwen to "upscale with model", use one of the Illustrious models for upscale - qwen now has a default template for that in case you even need one, just search for "upscale" in the template section.

 

[Midzay]

Qwen 2509 keeps growing on me. It is THE composition tool: you make a waify with Illustrious, then you put her thru the scenes with Qwen.

I spent quite a good teal of time with Illustrious and SD to know that Qwen brings an entirely new capability for pose control right out of the box. Granted, Qwen is one massive control next under the hood, but still.

Also check out the perspective with regard to full body shots and how natural it looks - this is new (at least to me).

View attachment 5313495	View attachment 5313499	View attachment 5313497
View attachment 5313544	View attachment 5313502	View attachment 5313503

FML, you are literally self-promoted to chief editor / senior photographer at Playboy circa 1985.

Make sure to employ an 8-step lora and "Patch Sage Attention" to cut down ~30% on generation time.

Also, don't use Qwen to "upscale with model", use one of the Illustrious models for upscale - qwen now has a default template for that in case you even need one, just search for "upscale" in the template section.

Explain to a regular Flux user in ComfyUI why they should use Illustrious first and then Qwen. Is this better than just generating in SD or Flux (with NSFW options)? Or is your limitation related to your video card?

 

[Sepheyer]

Explain to a regular Flux user in ComfyUI why they should use Illustrious first and then Qwen. Is this better than just generating in SD or Flux (with NSFW options)? Or is your limitation related to your video card?

Illustrious is fantastic for waifus. May be Flux too, I just dunno. SD in far inferior to Illustrious. So, in this context using Illustrious is to get the best looking girl - whatever that means for anyone - out there. Now, Qwen Image also has fantastic female faces and bodies, but I think the Illustrious is the king still -- there's way more checkpoints and the female anatomy is top notch.

Naturally, this is my opinion, so if anyone has solid reasons for why Flux should be used rather than Illustrious, I am cool with that, I trust you.

 

[Sharinel]

Illustrious is fantastic for waifus. May be Flux too, I just dunno. SD in far inferior to Illustrious. So, in this context using Illustrious is to get the best looking girl - whatever that means for anyone - out there. Now, Qwen Image also has fantastic female faces and bodies, but I think the Illustrious is the king still -- there's way more checkpoints and the female anatomy is top notch.

Naturally, this is my opinion, so if anyone has solid reasons for why Flux should be used rather than Illustrious, I am cool with that, I trust you.

Not sure I understand what you mean. Are you prompting an image with illustrious then using that image in img2img using Qwen? Or are you somehow refining with Qwen during the initial illustrious process?

 

[Sepheyer]

Not sure I understand what you mean. Are you prompting an image with illustrious then using that image in img2img using Qwen? Or are you somehow refining with Qwen during the initial illustrious process?

Yep, the former one.

 

[Sepheyer]

This is cray.

I have this one particular prompt that I kept testing - unsuccessfully - on the ControlNets going back to early SD. Namely I was trying to put yellow-tape marks on the floor and then put talent's feet on the marks.

Up to this point we were "not even in the same room".

And for the first time Qwen 2509 came thru. Just "boom!" and "here you go".

Quite an FML moment for me personally.

GOOD​	LESS GOOD​
... stands with legs spread apart wide, stands on two marks made of yellow tape crosses under each foot ...	... stands with legs spread apart wide, stands on yellow tape crosses marks under each foot ...

 

[Sharinel]

This is cray.

I have this one particular prompt that I kept testing - unsuccessfully - on the ControlNets going back to early SD. Namely I was trying to put yellow-tape marks on the floor and then put talent's feet on the marks.

Up to this point we were "not even in the same room".

And for the first time Qwen 2509 came thru. Just "boom!" and "here you go".

Quite an FML moment for me personally.

GOOD​	LESS GOOD​
... stands with legs spread apart wide, stands on two marks made of yellow tape crosses under each foot ...	... stands with legs spread apart wide, stands on yellow tape crosses marks under each foot ...
View attachment 5316877	View attachment 5316880

Any idea why everything I try comes up with a black screen? I'm using the inbuilt qwen image edit 2509 workflow and haven't changed the defaults, but no luck

 

[Sepheyer]

Any idea why everything I try comes up with a black screen? I'm using the inbuilt qwen image edit 2509 workflow and haven't changed the defaults, but no luck

View attachment 5317018

Nothing sticks out to me. Given you are no rookie in this, whatever is causing it probably one of those pain-in-the-ass things or so fkng obvious that we'll all have a laugh.

 

[JhonLui]

I'm not ComfyUi big fan, so I've seen there is a (fairly) new Inference called SD.Next that supports prettymuch everything: from Sd to Qwen, Flux and Wan... and it's actively updated.
Did anybody try it?

Also, your prompt gave me this... LOL
please give me other prompts!

 

[Sharinel]

Might be something to do with using it inside SwarmUI then

 

[JhonLui]

Might be something to do with using it inside SwarmUI then

Is CFG: 1 in the KSampler ok?
Swarm should't be the problem, it's litterally just a UI for Comfy

 

[Sepheyer]

Is CFG: 1 in the KSampler ok?
Swarm should't be the problem, it's litterally just a UI for Comfy

Yes, it is the 4 step lora that requires CFG of 1.

Sharinel check if you have all the models / loras in the places where the loaders think they are. Sometimes the warnings just don't come up.

 

[Sharinel]

Yeah I do, they all appear on the dropdown when I click to change the models/lora etc
Weird

 

[Sharinel]

Sepheyer JhonLui found out the cause! For some reason Qwen does not like --use-sage-attention as an argument when starting comfy.
Bit of a bummer that I need to swap it out all the time though if I want to use Qwen

 

[Sepheyer]

Sepheyer JhonLui found out the cause! For some reason Qwen does not like --use-sage-attention as an argument when starting comfy.
Bit of a bummer that I need to swap it out all the time though if I want to use Qwen

You prolly know that you can patch SA on a need-to basis using an SA node:



The CUI starts normally, completely unmolested, and if your workflow needs a SA then you just activate the node.

BUT! You can't deactivate it - need to restart.