SD3 - New License, Examples & Using it as a Refiner

Stable Diffusion 3 has a new license, so it looks like we're now good to go with commercial use... up to a limit. Do check out the information for yourself, available at - stability.ai/news/license-update
Here I do some basic prompt tests, compare outputs vs Pixart Sigma, try using SD3 as a refiner and also via high-res fix.
Want to support the channel?
/ nerdyrodent
/ sd3-licence-fix-107686733
* ComfyUI Extra Models - github.com/city96/ComfyUI_Ext...
== Learn more things! ==
1. Installing Anaconda for MS Windows Beginners - • Anaconda - Python Inst...
2. Installing ComfyUI for Beginners - • How to Install ComfyUI...
3. ComfyUI Workflows for Beginners - • ComfyUI Workflow Creat...
4. Pixart Sigma - • Pixart Sigma - Low VRA...

[@supercurioTube]

Thanks for showing so well what it can and can't do, beyond the already well understood anatomy challenges.

[@tad2021]

The small dataset safety training did before release caused many of issues, as did the apparently purging of a lot of the style captioning during training. From what I've been hearing, the other sizes SD3 were doing better, but SAI went released medium instead.

[@blackvx]

Mr. Rodent, thank you!

[@alexds9]

When comparing prompts, it's important to check multiple seeds to ensure the results are consistent and not an outlier. Checking 5 seeds for each prompt would provide a much more reliable and accurate representation of the model.

[@KadahCoba]

Try replacing clip L and G with trained ones from an SDXL model. SAI did not train clip for SD3, so its the same as SDXL. Results on SD3 can be better to varying extents depending on what SDXL model clip is taken from. You can either extract clip or grab is from a transformers version of a model and load it separately, or use a checkpoint loader and just pull clip from the node.
You can also use trained SD1.5 clip L models on SDXL, and maybe SD3 too. Though it we did such via differential merge of clip and applying it as a clip only lora. We found that for SDXL models with "less than ideal prompt control" (like trained with datasets with conflicting captioning), using a clip L lora from a better trained SD1.5 model had very good results.

[@TheDoomerBlox]

that one Mobius SDXL model has a pretty good CLIP model to use for this, if you want a general upgrade without locking yourself into a specific visual genre

[@marcfruchtman9473]

Glad to see the license improvements!

[@jibcot8541]

I have a workflow that is producing some great images for photographic , 20% SD3 noise for the good prompt aderance and 80% SDXL as a refiner.

[@tuna1867]

pls share it

[@TobinatorXXL]

Download? thanks 😀

[@samizdat_eth]

You always had the ability to make videos and commercials, even with the old license. This new one is a lot better overall though

[@abdelhakkhalil7684]

This is a half-cooked model shipped to the community to finish the job while StabilityAI reaps all the benefits. That's why they had that weird license before.

[@electronicmusicartcollective]

Thanx for this clear words

[@pon1]

Thanks!

[@sergetheijspartner2005]

I can't even draw a cow that good with crayons, so even the worst one is better then I can do by hand

[@sneedtube]

I'll wait for the true SD3 model (large? Idk) to drop before to consider giving it a shot for whatever purpose

[@zephilde]

Didn't you tryed the other way? SD3 for text & composition/comprehenssion, and PixArt for style refiner/hires , to get the best of both worlds ;)

[@juanjesusligero391]

As always, splendid video! Thanks for sharing! ^^ Are you keeping tabs on the 'Open Model Initiative'? They're cooking up some truly free models! :D

[@SouthbayCreations]

Howdy! Thanks for the video, great info! Did you post your workflow anywhere? Thank you!

[@vi6ddarkking]

The fact SD3 sucks by default is fine. It's a base model.
SD1.5 and SDXL also left much to be desired.
Now with the fine tunes and the SD3 refresh they are working on.
As long as we have the excellent prompt understanding the community can improve it greatly over time.

[@tazztone]

so are good sd3 models popping up on civitai any moment ? :) lookin forward

[@Elwaves2925]

@@tazztone Last time I checked they hadn't removed their ban of SD3. My guess is they're still running the new license by their lawers....just to be sure.

[@97BuckeyeGuy]

Looks like it's up to the community to train SD3 into "not suck" territory, now.

[@AkshayKumar-vd5wn]

@@97BuckeyeGuy Likely.

[@sergetheijspartner2005]

It's a Uniroo, mix between a unicorn and a Kangaroo

[@luke2642]

Surely you got it backwards...SD3 for prompt adherence of the base image, PIXMA as a refinery/hires?

[@havemoney]

It’s as if Civit hasn’t lifted the ban, a day has passed and there’s not a single new lora (

[@AkshayKumar-vd5wn]

Still waiting for a free alternative that beats Midjourney.

[@ameet21]

If they perfected the model... only then make sense to invest money for licence. I mean for now I feel it's wastage only on tests which you can do it with sdxl or sd1.5 even online image generation who offer best with prices. What's the point of wasting money when we don't get desire effects and look and live with crap with licence which you know you can't use commercially

[@ayrengreber5738]

This new license is only a step away from the old one commercially. They still hold power to revoke your commercial access to the model.

[@Elwaves2925]

That's the case with almost any license anywhere. Having a license doesn't mean you get it indefinitely unless you specifically pay for that.

[@Elwaves2925]

@@AB-wf8ek I said pretty much the same thing with the original license. Stability have no way to legally check what you create offline or locally while online. You can also strip the data from the generated images so they can't check that after the fact either.

[@sherpya]

they need to release a fixed model first

[@Ethan_Fel]

SD3 is pretty much dead, the license correction came too late and they lost too many talents

[@fast_harmonic_psychedelic]

Our loss.

[@juanjesusligero391]

@@fast_harmonic_psychedelic Yeah, it's a shame. But luckily there is also a bright side, since some of those people joined together and are making new models, wich will be 100% FOSS :D (they are thinking of using Apache or MIT license for the models)

[@Ethan_Fel]

@@AB-wf8ek nobody or nearly nobody is finetuning sd3 though.

[@bossmassful]

@@AB-wf8ek i remember how i trying sdxl base for the first time. Hilarious piece of... After fine tuned 1.5 models. Unbelievable.
But now... Deis sampler + AYS scheduler, advanced workflows. 2 merged sdxl models, fine tuned of course. Loras sometimes. Enough for me.

[@Ethan_Fel]

@@AB-wf8ek you're coping. Adults take into account the cost and the potential of anything they do. It's quite obvious that SD3 isn't it for the people able to finetune.
Especially since the revised license has still some glaring issues

[@GreenAppelPie]

If SD3 was useful, it might matter.

[@AIwaysUploading]

Why are you even here lol

[@contentfreeGPT5-py6uv]

Sd3 Is the best free for me,every 1000 images with 0$

[@CharlesLijt]

Stop using this shit, Kolors is the new King