A relatively successful, if I didn't say so myself, antique color photography LoRA for FLUX.1-schnell, improving it more reliably than most Dev-trained LoRAs.
Plus, as Schnell goes, sell whatever you might, and fear not the legal eagle.
Catered to producing realistic images, especially portraits, reminiscent of color film analog photography, exhibiting parallels to a broad spectrum of iconic instrumentalities and visual paradigms, from Autochrome-to-Kodachrome-to-Fujifilm-and-beyond.
Produces visuals with a vaguely "historical" or "lived-in" aesthetic character, striking chromaticity and luminosity dynamics, as well as richer textural/anatomical/skin details.
Trained by me, A.C.T. Soon®, for 4000 steps at full rank 64 on one A100 via Colab Pro, using the dedicated Schnell Training Notebook by Ostris on 135 color photographs taken during the 1900s and 1910s by Sergey Prokudin-Gorsky, who traveled and photographed widely in those years while pioneering and perfecting implementations of an early three-color-composite photography technique.
We urge you to explore the work of Prokudin-Gorsky for yourself, at the wonderfully organized online archive at this link, featuring many hundreds of high quality downloadable scans of composite color photo prints from the photographer's original glass plate negatives, available at this site alongside an archive of recent restorations and more. The original Prokudin-Gorsky glass-plate negatives are currently held at and administrated by the Library of Congress in Washington, DC, USA, along with a collection of prints.
You should use the token HST to trigger the image generation.
Historical Note:
Prokudin-Gorsky's color photography technique would involve three photo-exposures, either simultaneous or sequential, using specialized color-spectrum filters (basically R.B.G.: red, blue, and green), rendering a subject/shot onto glass plates covered with light-emulsive mixture.
The photographer's focus on refining the developer and filter quality, in tandem with his incessant and wide-ranging experimentation, and his artful optimizations of glass plates (generally unwieldy, esp. for color, and by the 1910's already becoming outmoded for B&W on-location shoots, though elsewise extra reliable) ultimately led him to produce a color photography oeuvre of much greater fidelity and vividness than achieved by most of his contemporaries.
At the same time, the peculiarities of the photographer's method, coupled with his exceptionally hands-on execution thereof, would manifest in a range of idiosyncratic color, light, and motion artifacts common across the resulting prints.
Seldom marring the image as a whole, and less grave than the weaknesses of some contemporaneously emerging autochrome techniques, the warm color hazes and flares framing many of Prokudin-Gorsky's prints may be seen as a kind of ephemeral signature.
Alongside some of the more subtle chromatic, textural, and (in some measure) figural characteristics of his work, these auras have reliably imprinted themselves into this Flux Schnell LoRA.
Description
Converted to the sd-scripts LoRA format using Kohya's conversion script. It's remarkable how much of the transformer layers structure has to get "translated" between the two formats; yet the architecture remains, more or less, analogous.
FAQ
Comments (4)
It seems that Schnell trainings goes towards dataset with more "impact" than DEV ones but still it's a less powerful model on the inference. Question... which one has the accurate aesthetics in the training? Kohya or AIToolkit? Im still wondering about that,
I actually have only trained on ai-trainer, and have converted to Kohya mainly to use the sd-scripts LoRA merger scripts to experiment with co-merging into each other LoRAs I've trained on same or similar datasets (under different settings + over different base layers), towards a broader WIP process of trying to alchemize a custom checkpoint. And I've uploaded both versions here sheerly to provide broader inference capability, but the inference behavior should be fairly similar. There are some differences however. And I appreciate you bringing this up, because I didn't actually bother to do a true comparison of the versions before uploading, but I just did so now, and will post it in a minute into a gallery. As I expected, the differences don't seem pronounced or specific enough thus far to categorically proclaim one or the other version "superior" in a non-ambiguous way. But maybe others would discern more obvious differences.
And as far as Schnell vs. Dev: Broadly speaking, I agree with the gist of what you're saying. However, the following factor still remains in play: Even though it proved relatively simple to somewhat placate Dev's steeper hardware/step requirements, no amount of fine-tuning, training, or any other modifications would enable anyone to outrun its more restrictive license.
Therefore, Dev is only open source tenuously, and any model DEVeloped on its basis could potentially be at any point reclaimed or shut down by whoever owns (at that point in time) the rights to the base Dev (and there's no guarantee that would always be the same people). And even distribution privileges could technically be revoked at their whim, and under any premise, and that would still be legally sound.
Granted, and this is something that many don't seem to get, this restrictive non-commercial license and its implications only extend to models, and not to inference outputs, whether they were generated with Dev, Schnell, or Pro.
But IMHO the potential value of modifying Schnell, and the tangible value of refining inference results, ought not be evaluated as compared against Dev, but only in comparison with base Schnell, since it is the only open source Flux model we actually have any firm collective claim to.
P.S. I uploaded comparisons into the gallery here.
Yeah, I still didn't test Schnell on Kohya cause require a bit more of attention and time than I just have right now but at least and from a very wide angle I can saw that Kohya's training Schnell dragged way moure the input (as if XL training) than AI Toolkit's training on DEV because it was more "adaptative". Was interesting to see such behaviour.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.