[FRIAM] Dabble with Stable Diffusion and like UV. and have questions

[glen]

I haven't had good luck with local VLMs. Granted I've only tried a few and don't keep track of what I've tried. My last one was Z-Image <https://github.com/Tongyi-MAI/Z-Image>, in January. I'm clearly not good at this. I'm not posting the other prompts because they are NSFW. But it should be clear the only one that was successful was the Irish Judo prompt:

"Red-haired girl in a black Judo gee throwing a roundhouse kick. Irish countryside background. Slight rain. Near dusk lighting. Moon at slight angle from horizon."

Too many failures to list in the South Park one ... though I was impressed they got the words in the bubble almost right. Try as I might, I couldn't get the dagger in the Satan Trump one to go *into* the baby. Renee' pays the electric bill. I need to get the report for January. 🥺

p.s. Sorry for the yuge photo. I used jpg in this one.

On 3/23/26 5:10 PM, Steve Smith wrote:
> I could be missing something... it feels as if the generated image is *yet more* similar to the one you seeded with than I would expect.  It is probably spurious that you did not include a harmonica (from the original) in the generated one.    Maybe the "crossed drumsticks" provided a better perceptual/conceptual anchor than I would expect?
> 
> Probably not coincidental that the specifics of each element was so "generic"... I'd have thought there was more variaiton in style and sizing of the drumsticks, recorder, tin-whistle, etc.   Even the "iron" tuning forks were more "normalized" than I would have expected.   The originals, while probably steel (stainless), were surprising close in scale/design than I would have suspected.
> 
> This all says much more about my expectations and how "normalized' these types of instruments/elements would be.
> 
> The *patina* on the objects (sans recorder?) feels acutely authentic excepting the pentagram box whose patina seems to be simply "wood-charring" of a modern-factory-constructed box with little to no wear?   I'm trusting that the ?runes? on the margins of the box are "authentic", but might have been drawn from some more nominal/stylized portion of the plenum?
> 
> Re: energy/water-cooling consumption:  I wonder how much a (more) full distribution of transformers (from cloud to edge) will shift that?  When my laptop/phone is doing a significant part of the work, (or a federation of semi-idle laptops/phones) will the waste heat/entropy-bump change in any way or just occur in different geolocations?
> 
> On 3/23/26 10:34 am, glen wrote:
>> Nice. I've been playing with "round trip" variation. I'll send an image to an LLM (Claude, Kimi, Mistral, Gemini, etc.) and have them describe it, then either generate variants directly (for the ones that can do it) or explicitly ask for a prompt to send to a visual model - Claude can't generate images, yet). Then I'll take their generated image(s) and send them to one of the others and ask if that image is AI generated. I can't help but wonder how much fresh water and electricity I'm wasting do this sort of thing; c'est la vie.
>>
>> Attached is one of my original pictures and a variant generated from Claude's prompt. It was interesting that I uploaded that same photo to several LLMs and *only* Claude named the Jaw Harp. All the others said things like "a key-like tool of some kind". Here's one of Claude's prompt variations:
>> "Top-down view of ancient musical instruments laid out on a dark wooden altar — a bone-white recorder, a tarnished tin whistle, crossed drumsticks, iron tuning forks, and a jaw harp — beside a glowing carved pentagram box. Candlelight flickers. Smoke wisps. Dark fantasy RPG atmosphere."
>>
>> On 3/23/26 9:06 AM, cody dooderson wrote:
>>> Hi Gillian and the FRIAM Group,
>>>
>>> I am going to thread jack because I know nothing about pip or UV, but you did mention stable diffusion. I experimented with the HuggingFace Diffusers library a few months ago(using pip).
>>>
>>> My favorite thing to do was make the models hallucinate. Below is one of my favorite results featuring a coworker; I was trying to put a person into the style of a photograph using the Stable Diffusion IP-Adapter pipeline. When I set the guidance for style to 0 and the face to 1, it created all sorts of crazy hallucinations. I have don't why it kept putting sheep into the picture. I could probably have an art show with all of the failed stable diffusion images I made.
>>>
>>> Thanks for opening up the window for me to share these!
>>>
>>> xu.png
>>> Different levels of hallucination:
>>> diffusion10-20step.png
>>>
>>>
>>>
>>>
>>>
>>> On Mon, Mar 23, 2026 at 8:42 AM glen <gepropella at gmail.com <mailto:gepropella at gmail.com>> wrote:
>>>
>>>     I'm sure others are more well-informed than me. But because I'm partly responsible for the overwhelming number of tokens in the other thread, I feel a responsibility to reinforce your post.
>>>
>>>     My guess would be UV (being in Rust) is more naturally parallel than PIP. So parallel downloads, parallel caching, etc. But to exploit that, you'd have to also have a good way to manage package dependencies. E.g. if any 2 packages, and the dependency trees they're in, can be proven independent, than those trees can be downloaded and installed in parallel.
>>>
>>>     But, as I said, others may have better or more detailed ideas.
>>>
>>>     On 3/21/26 10:08 AM, Gillian Densmore wrote:
>>>      > Basically dabbling with Stable Diffusion.
>>>      > One of the webuis for it uses UV.
>>>      > It seems pretty ok. I am confused though: How is that pip has latency and download speed issues etc. While UV seems to go faster. The command lines being readable by this idiot helps.
>>>      >
>>>      > Some mine doing a ELi5 what black magic UV's devs do to get it a little better optimized. Seems to resolve deps nicer as well...I tried reading on git hub how Astral (or other devs?) can do that. And got lost.
>>>      >
-- 
¡sıɹƎ ןıɐH ⊥ ɐןןǝdoɹ ǝ uǝןƃ
ὅτε oi μὲν ἄλλοι κύνες τοὺς ἐχϑροὺς δάκνουσιν, ἐγὰ δὲ τοὺς φίλους, ἵνα σώσω.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: z-image-montage.jpg
Type: image/jpeg
Size: 348524 bytes
Desc: not available
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20260324/b4763867/attachment-0001.jpg>