As long as they have a version that supports it, you can flip over to the Web UI to set it, then go back to a native client, if need be.
Kbin/Mbin have supported instance blocking since forever. Dunno about piefed and sublinks.
tal
joined 1 year ago
Yeah, user migration would be nice.
If it were a shift to simply using a keypair as the basis for identity, which would be a big change, then one could potentially transparently use any instance. That'd be neat from an instance reliability standpoint.
Keypair-based identity would also permit migrating an account from a permanently failed instance. Right now, the home instance is the authoritative source for the account. The problem with that is that if the instance goes away forever, then there's no authoritative source left to determine who controls a user account. One of the use cases that I'm worried about is a big instance going down because the admins get in a car crash or something, and it killing all the user reputation that's been built up, because nothing can be done after the permanent failure.
IIRC feddit.uk had a close call like this a while back.
What would that involve? I mean, you can already have spoiler sections in post body text.
Lemmy.today is in Oregon.
Honestly, though, as long as your home instance is in the same half of the US, or for Europe, anywhere in Europe, I doubt that latency is gonna be very noticable.
That helps, but honestly, it shouldn't be a client workaround.
Detect AI prompts attached to images and not strip them from images on upload the way other EXIF data is.
Stripping EXIF location data to help keep people from being doxxed is one thing, but AI image generators try to make images with metadata to indicate that the images are AI generated, which helps avoid using them for training and lets people inspect how images are created. As of now, that gets stripped on upload. It's particularly obnoxious over in !imageai@sh.itjust.works.
EDIT: Even nicer would be the option to leave EXIF location data attached, and merely warn a user at upload time about location data and provide the option to strip it, as I can certainly imagine communities where people would really like to be able to include precise location data with their images.
I haven't looked into any actual decision process, and personally, I'd like to be able to post vector files myself, but there are some real concerns that I suspect apply (and are probably also why other sites, like Reddit, don't provide SVG support).
-
SVG can contain Javascript, which can introduce security concerns
-
My guess is that there are resource-exhaustion issues. With a raster format, like, say, PNG, you're probably only going to create issues with very large images, which are easy to filter out -- and the existing system does place limits on that. With a typical, unconstrained vector format like Postscript or SVG, you can probably require an arbitrarily long amount of rendering time from the renderer (and maybe memory usage, dunno, don't know how current renderers work).
-
At least some SVG renderers support reference of external files. That could permit, to some degree, deanonymizing people who view them -- like, a commenter could see which IP addresses are viewing a comment. That's actually probably a more general privacy issue with Lemmy today, but it can at least theoretically be addressed by modifying Threadiverse server software to rewrite comments and by propagating and caching images, whereas SVG support would bake external reference support in.
I think that an environment that permits arbitrary vector files to be posted would probably require some kind of more-constrained format, or at minimum, some kind of hardened filter on the instance side to block or process images to convert them into a form acceptable for mass distribution by anonymous users.
Note that Lemmy does have support for other format than SVG, including video files -- just not anything vector ATM.
If the art you want to post is flat-color, my guess is that your closest bet is probably posting a raster version of it as PNG or maybe lossless webp.
Can also store an SVG somewhere else that permits hosting SVG and provide an external link to that SVG file.
There has to be soy milk or almond milk or something else that could be used.
If you have a gun, never get within knife range.
breathtaking cat in neon paints . award-winning, professional, highly detailed
Negative prompt: signature, text, watermark, ugly, deformed, noisy, blurry, distorted, grainy
Steps: 20, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 3, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.16, Token merging ratio: 0.5, Token merging ratio hr: 0.5, Hires upscale: 2, Hires upscaler: R-ESRGAN 4x+, Version: v1.7.0-133-gde03882d
city, diamine art, shimmering
Negative prompt: bottle, photograph, text, signature
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 0, Size: 2560x1440, Model hash: ebf42d1fae, Model: realmixXL_v15, Version: v1.7.0-133-gde03882d
A couple notes:
-
I'm aiming for a fountain pen look; you can do quite a bit with fountain pen ink, creates a lot of color gradations and so forth, though in real life, the ink is hard to control. I really like the look of the stuff. It's kind of like more-elaborate watercolors. I'd spent some time in the past unsuccessfully trying to get such a look with "fountain pen" and similar terms, and didn't get satisfactory output; I got pictures of fountain pens, but not much that looked like a fountain pen artwork. I finally hit it by trying specific ink names; the one here is reference to a line of a "shimmering" inks made by Diamine.
-
This image was generated natively at 2560x1440; apparently, at least with
--medvram, this is possible on a 24GB video card. Automatic1111 does not, by default, permit a user to create images larger than 2048 in any dimension; typically, users upscale to these resolutions. However, one can editui-config.jsondirectly and modifytxt2img/Width/maximumto be higher numbers and it will work, as long as there is enough video memory. -
Stable Diffusion tends not to do so well generating images much larger than the training size; what I expect happens is that it starts to converge on different images in different parts of the large image, and doesn't wind up having the image as a whole converge. I would guess that it's possible to tweak the ancestral noise settings so that there's enough noise added at each stage to bump it out of whatever local minimum it's converged on, but at least with the standard settings, this isn't really possible. This means that one tends to get the sort of "distorted monster" look with lots of people merging into one. I ran through a couple different types of scenes, looking for something that wasn't too-badly impacted; I'd noticed before that landscapes tended not to be too badly impacted, as Stable Diffusion could reasonably fill in, say, a cliff face between two existing cliff faces that have been converged on in a way that it can't fill in when two different human faces that collide with each other have been converged on in different parts of the image. Cityscapes also seem to do all right; SD can fill in similar buildings, fit things together pretty well. Basically, one wants a scene that doesn't have giant features that can't reasonably be reconciled with each other.
An image of Santa Claus and Krampus.
This was a two-step generation. First, creation of a 1280x720 image. Three columns were used in Regional Prompter; one column in the center that used both Santa and Krampus as a negative prompt to maintain some separation between Santa and Krampus in the image:
gouache illustration
ADDCOMM
santa claus
ADDCOL
ADDCOL
krampus
Negative prompt: woman
ADDCOMM
ADDCOL
santa claus, krampus
ADDCOL
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1, Size: 1280x768, Model hash: ebf42d1fae, Model: realmixXL_v15, RP Active: True, RP Divide mode: Matrix, RP Matrix submode: Columns, RP Mask submode: Mask, RP Prompt submode: Prompt, RP Calc Mode: Attention, RP Ratios: "1,.5,1", RP Base Ratios: 0.2, RP Use Base: False, RP Use Common: True, RP Use Ncommon: True, RP Options: [False], RP LoRA Neg Te Ratios: 0, RP LoRA Neg U Ratios: 0, RP threshold: 0.4, RP LoRA Stop Step: 0, RP LoRA Hires Stop Step: 0, RP Flip: False, Version: v1.7.0-133-gde03882d
The Hires fix upscaler built into Stable Diffusion doesn't work with a regional prompt, as each tile it works on will have the regional prompt applied to it. So it's not possible to generate a full-resolution image in a single step.
I took the image and 2x upscaled it to 2560x1440 as a separate step in img2img using the Ultimate SD Upscaler extension:
gouache illustration
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 824048783, Size: 2560x1536, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.16, Ultimate SD upscale upscaler: SwinIR 4x, Ultimate SD upscale tile_width: 512, Ultimate SD upscale tile_height: 512, Ultimate SD upscale mask_blur: 8, Ultimate SD upscale padding: 32, Version: v1.7.0-133-gde03882d
That produced the final image.
rooster, neon paints, black background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 15, Size: 1024x1024, Model hash: ebf42d1fae, Model: realmixXL_v15, Version: v1.7.0-133-gde03882d
dia de los muertos, sailor moon
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 7, Size: 1024x1024, Model hash: ebf42d1fae, Model: realmixXL_v15, Version: v1.7.0-133-gde03882d
A series of famous present-day landscapes in the US picked from this list rendered in the style of John William Casilear, a 19th-century American landscape painter. They don't look quite like the actual locations -- to do that, I expect that one would probably do better to start with a photograph and do img2img transforms -- though they do get new, interesting landscapes with the feel of them.
Rendered in Stable Diffusion.
These are 2560x1440, aiming for a full-screen image on a 1440p screen, and are probably suitable for use as desktop backgrounds on such screens. If you are viewing these on a desktop in 16:9 aspect ratio (the most-common these days), you can probably benefit from opening the images and viewing them fullscreen. I include a link for each, as the lemmy Web UI doesn't have a great way for a user to view an enlarged inline image (IIRC the kbin Web UI does).
Scotts Bluff, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 12, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Three Sisters Springs, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 6, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Kent Falls, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 19, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Avenue of the Giants, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 18, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Buffalo National River, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 10, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Denali, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 15, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Dismals Canyon, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 13, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
Santa Elena Canyon, by John William Casilear
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2, Size: 1280x720, Model hash: ebf42d1fae, Model: realmixXL_v15, Denoising strength: 0.7, Hires upscale: 2, Hires steps: 10, Hires upscaler: Latent, Version: v1.7.0-133-gde03882d
-
I usually use the Euler a sampler; I find that the ancestral samplers ("-a") tend to do better on things like fingers (though they have the technical drawback that adding more samples also alters the image; you can't just "converge" on a single image and throw the minimum number of samples required at the problem). For at least these painted landscapes, I didn't seem to have any visible issues using a non-ancestral sampler.
-
I tend to render at 1024x1024, the SDXL training resolution; this helps to avoid things like people with extra limbs. For landscapes, this doesn't seem to be an issue for me, and using a different aspect ratio seems to work without flagrant visible issues. My experience is that it can lead to similar elements replicated in an image, but doesn't seem to be a crippling issue for these landscapes.
-
One of the largest issues I ran into was reflections in water that didn't quite match the image above the water; I tended to select images without reflections in the water, or at least ones where the issue was less-obvious, though having reflections in water is common in landscape paintings and in Casilear's actual works. My impression is that Stable Diffusion can deal well with situations where a perfect mirror of an image is expected, but not with one that is distorted in various ways -- as is the case for water with ripples.
-
I usually post images in PNG format. I use JPEG here; they are less than a tenth the size for images in this style, and I feel like doing PNG for the larger images I'm doing here places unreasonable load on the lemmy host I use, which is -- at no charge -- hosting posted images for users of the server. I wasn't able to see any artifacts in visually-inspecting PNG and JPEG versions.
-
I rendered batches of 20, and picked the output that I subjectively felt was most-aesthetically-pleasing.
-
Prompts and other information to reproduce the images are included for each.
I mean, it's doable now, but I think that the limiting factor is just the userbase. More developers using the platform, more people interested in writing code for browser extensions.
There is a lemmy/kbin assistant extension for Firefox, which is far, far more basic than RES, but provides one critical feature that I regularly use -- being able to view a post on one's home instance. So people have done work on these.
Also, if by "Old Lemmy", you mean mlmym, that's not merely the website. It's an alternate Web UI that instances can run alongside the regular one. My home instance does so at https://old.lemmy.today/
EDIT: Your home instance does as well, at https://old.lemmy.world/