PixelWave FLUX.1-schnell 04 - Apache 2.0!
Safetensor Files: 💾BF16 💾FP8 💾bnb FP4
GGUF Files: 💾Q8_0 🤗Q6_K 💾Q4_K_M
Model also available at: RunDiffusion and Runware.ai
PixelWave FLUX.1 schnell version 04 is an aesthetic fine tune of FLUX.1-schnell. The training images were hand picked to ensure the model has a bias to eye catching images, with beautiful colors, textures and lighting.
Trained on the original schnell model, so Apache 2.0 license!
No special requirements to run. Supports FLUX LoRAs
Euler Normal, 8 steps.
You can use more steps to improve finer details, but the output doesn't change much after 8 steps.
Shout out to RunDiffusion
Huge thank you to RunDiffusion (co-creators of Juggernaut) for sponsoring the compute that made training this model possible! Figuring out how to train schnell without de-distilling the model required a lot of experimenting, and being able to utilize RunDiffusion's cloud compute made it a lot easier.
For those needing API access for this model, we're partnering with Runware.ai
I have made the FLUX.1-dev 04 version exclusive to RunDiffusion and Runware for the time being. When I release version 05 in future, I plan to release the dev 04 open weights.
Grateful for their support in getting this model out there, please check them out!
Training
Training was done with kohya_ss/sd-scripts. You can find my fork of Kohya here , which also contains changes to the sd-scripts submodule, make sure you clone both.
Use the fine tuning tab. I found the best results with the pagedlion8bit optimizer which also could run on my 4090 GPU 24GB. I found other optimizers struggle to learn anything.
I have frozen the time_in, vector_in and mod/modulation parameters. This stops the 'de-distillation'.
I avoid training single blocks over 15. You can set which blocks to train in the FLUX section.
LR 5e-6 trains fast, but you have to stop after a few thousand steps as it starts to corrupt blocks and slow down learning.
You can then block merge with an earlier checkpoint, replacing the corrupt blocks, and then continue training further.
Signs of corrupt blocks: paper texture over most images, loss of background details.
Contact
For business or commercial inquiries please reach out to us at [email protected]. Licensing flux fine tunes. Customer training projects. Commercial AI development. The team can do it all!
PixelWave Flux.1-dev 03 fine tuned!
Safetensor Files: 💾BF16 💾FP8 💾NF4
GGUF Files: 💾Q8_0 🤗Q6_K 💾Q4_K_M
The 'diffusers' files are actually the Q8_0 and Q4_K_M GGUF versions. GGUF files also available on huggingface.
I fine tuned version 03 from base FLUX.1-dev for over 5 weeks on my 4090. It is able to do different art styles, photography, and anime. Trick I discovered to help with LoRAs.
I used dpmpp 2m sgm uniform 30 steps for the showcase images. If you want a neater/cleaner output, try increasing the guidance. Also mentioning a style can help, so the model doesn't have to guess.
I also recommend try adding the upscale latent by node, and scale the latent by 1.5, e.g. generating an image that is 1536x1536 instead of 1024x1024.
PixelWave Flux.1-schnell 03
GGUF Files: go to huggingface
I used dpmpp 2m sgm uniform 8 steps for the showcase images.
You can start with 4 steps, but there are less errors with anatomy if you run with more steps.
PixelWave Flux.1-dev 02
GGUF Files: 💾Q8_0 🤗Q6_K 💾Q4_K_M
Version 02 has greatly improved black and dark images, and more reliable outputs with fewer issues with hands.
I recommend using dpmpp_2s_ancestral, beta, 14 steps. Or euler, simple, 20 steps.
PixelWave 11 SDXL. A general purpose fine tuned model. Great for art and photo styles.
I use 20 steps, DPM++ SDE, CFG 4 to 6 or 40 steps, 2M SDE Karras
Accelerated Version - 5+ Steps, DPM++ SDE Karras, 2.5 CFG
PAG Recommended⚡Recommend 1.5 Scale, with CFG 3. Link to workflow
⭐Link to prompting guide.⭐ You don't need to use 'quality' terms such as 4K, 8K, masterpiece, high def, high quality, etc. Unless you want it, I recommend not using words such as 'vibrant, intense, bright, high contrast, neon, dramatic' for photographic styles if you a wanting a more natural look. This can cause images to look 'overcooked', but it's just the CLIP following your prompt. 🙂 If you do want vibrant, neon photos PixelWave will provide!
The focus for version 10 was to train the CLIP models, which improves the reliability, ensures you can produce a wide variety of styles, and better at following prompts.
Thanks to my friends who helped test: masslevel, blink, socalguitarist, klinter, wizard whitebeard.
Guide: Upscaling Prompts with LM Studio and Mikey Nodes
Guide: Add more details to your image using the skip step method
No need for the refiner model.
This model is not a mix of other models.
I also created Mikey Nodes which contains a lot of useful nodes. You can install it through comfy manager.
Description
Large 190k image LoRA on the double blocks as a kind of pretrain. Then multiple LoRAs trained and added to the single layer blocks to introduce art styles and improve finer details like skin textures.
The black levels and dark images have been greatly improved in this version. I've been very careful to not screw up the hands too.
FAQ
Comments (35)
any chance of doing a Q6 for the new version?
Well if you read the first paragraph he said: "I'll upload the rest of the GGUF versions here too when the site supports it." lol
When I made version 01, I made a lot of different versions. But I found Q4_K_M and Q8_0 were the only versions that didn't screw up the hands in my tests.
But I have uploaded the Q6 gguf to HF for you: https://huggingface.co/mikeyandfriends/PixelWave_FLUX.1-dev_02/blob/main/pixelwave_flux1_dev_Q6_K_02.gguf
@humblemikey thanks mate your the best. cheers!
@colinw2292823 im not talking about uploading it here, not sure why you feel the need to comment. im not talking to you
Awesome model! Just by any chance, do you know of any guide on how to merge loras into Flux? I've trained some but they're not working at all after the merge
Suggestion: Maybe try to further merge your Schnell checkpoint with hugovntr's flux-schnell-realism LoRA (https://huggingface.co/hugovntr/flux-schnell-realism) and my Historic Color Schnell LoRA (fine-tuning Schnell on high quality early color (not colorized) Prokudin-Gorsky photographs from around 1910, with a surprisingly decent outcome) at https://huggingface.co/AlekseyCalvin/historic_color_schnell . I've been getting wonderful results with this combo (using your v1 Schnell PixelWave merge with these 2 LoRAs: adding schnell-realism at between 0.6 to 0.8 scale and Historic Color at 1.0 or between 0.8 and 1.2). Just posted a basic 4-step cat into gallery here as an example. Might add a few more in a minute.
Thanks for the suggestion. But I intend PixelWave series to be it's own thing with no merges and avoiding synthetic AI images for training data. This doesn't mean you can't make a merged model yourself and post it. That would be awesome! Please share here if you decide to do this. 🙂
That makes sense. And I might take you up on that, especially if I set up a Flux merging workflow locally, despite my low-tier hardware. But if I get it to work, I'll certainly share the results here. And by the way, I totally support you on the no-synthetic-data in training principle. I am frankly a bit confused by how the use of "synthetic data" is somehow again becoming seen as a sensible and even laudable training/fine-tuning practice. Whatever happened to all that concern floated for a while about model corruption/banalization from overfitting, hypernormalization, and all the rest of it?!
But in any case, your stated aims and training practices/stipulations now compel me to offer an alternate suggestion. https://prokudin-gorsky.org/gallery/tag Through this link lives a very special archive of old photographs, one which even a cursory browse should confirm as, vitally, a visual document vein of a flabbergasting uniqueness and beauty. Contained therein are high definition detailed full color (not colorized) photographs from between 1905 and 1920 or so. I already mentioned it on a side note in my first comment, as one of my own current top sources for fine-tuning image sets, but now I just want to share it as a resources. After all, these photos are about as far as anything gets, in both origin and character, from synthetic data or content. To this, one may rightly wonder: how could any visual document or content be more "organic" or more "human-sourced" than any other, (unless in mere binary contrast with the purely synthetic...)?! Nonetheless, I'd insist: if ever there was such a thing as arch-organic photography, then this archive would be among its prime representations.
And if you do check it out: note that most of the photographs in this archive appear therein in several versions. Typically, between one and several alternate restored versions, plus the "original" scan of Sergey Prokudin-Gorsky's o.g. photo negative(s) (with each image being a composite of three photo negatives, shot either sequentially or simultaneously, and each lens-filtered to capture a dedicated partial segment of the color spectrum: basically an early iteration of the R.B.G. or Red, Blue, Green principle). These originals all share a distinctive frame, and a kind of glowy warm chromatic quality. Many of them also exhibit analog color and motion artifacts, mostly extending from the somewhat clunky hands-on logistics of the unusual, both at the time (1910s) and to this day, photography technique (though fairly similar to run-of-the-mill autochrome early color photography, in some select ways Prokudin-Gorsky went about it like a real provincial Eastern European weirdo, stumbling through to unparalleled and downright implausible results). Ultimately, out of the archive selections, though the originals are more unique, the restored versions might be, by and large, more useable towards generalized training sets.
Wow, I seem to have written a lot more than intended. Truth is, I've just been in utter awe of this archive ever since discovering its existence about a week ago, and wanting to share this discovery with anyone and everyone who might plausibly or even remotely care. These photos are from the 1900s and the 1910s, yet many of them really look like they could have been taken yesterday! Except their contents are the literal non-synthetic Russian Empire and places/views in central/Western Europe in the literal 1900s and 1910s. And there are many many hundreds of photos.
@alekseycalvin303 Thanks for sharing that resource! I'll have a good look through it.
Better than vanilla flux in a lot of ways, realism is up, although it lost some of its ability to spit out a variety of faces by default without specific prompts, There's definitely a face that ended up averaged into this model.
Thanks for the feedback. Definitely something to work on. I had a LoRA trained on a few thousand different faces, but the skin texture ended up looking rubbery and not convincing.
would love to see a full version of this fp16.
I have uploaded a fp16 version.
Thanks!
Summary: this checkpoint isn't better or worse than base Dev, it's just different. You do loose some style and prompt adherence.
For the people wondering how it compares to base for style, prompt adherence, anatomy, etc (variety of prompts, subjects, etc):
400 images
30% were kind of better
60% similar
10% worse than base
Biggest difference is the style. It's more like unsplash/modern film imitation/lomo/etc.
Hi, I’m hoping you can help clarify something for me. Yesterday, I downloaded the 11GB version of Pixelwave_Flux1Dev02, and it worked perfectly. Today, I downloaded the 22GB update with the same name, but now everything is pixelated no matter what I try. Am I missing something here? Using Forge...
Same issue. Using FP16 settings and recommended Sampler in SwarmUI. The image is fine with no Loras. If I add loras, the images are pixelated and messy.
Compared to the first version, the realism has noticeably worsened! I'm upset (Forge) Go back to the first version!!!
lol who are you to be upset, he is working and sharing it freely calm down
Not to worry, you can always use the first version - it's not going anywhere. I still use the old Pixewave version 6 for SDXL for realistic portraits ;)
Getting some great results with this. Any chance for Schnell GGUF? TYSM!
Under the files section, the files labeled as diffusers are actually the GGUF files I zipped. Civitai currently doesn't support uploading GGUF files. You just need to download and unzip.
@humblemikey Ah I found it ty!
This is the best realistic Flux model I've come across so far. I'm mainly using Flux in an img2img workflow with source images generated using Pony models. This model gives the best realism, skin details, as well as detailed backgrounds, unlike other popular models where everything end up smoother and/or washed out. But maybe that's just me...
Looking forward to future iterations of this great model!
Version one or version two? For me, version one gives more realism.
Amazing work! Version 1 looks better for realistic images, skin texture is amazing, but v2 is better for creativity focused on artistic, phantasy and illustration. Maybe you should consider creating two versions of pixelwave. Thanks!
Just FYI: I performed a conversion of the fp8 Schnell v1 checkpoint into the Huggingface/Diffusers format and made a repo/repost of it at HF, with all the due credit/links back here:
https://huggingface.co/AlekseyCalvin/PixelwaveFluxSchnell_Diffusers
I also re-posted the Schnell Safetensors there in a separate repo:
https://huggingface.co/AlekseyCalvin/PixelWave_Schnell_v1_fp8_safetensors_by_humblemikey
I made these repos (particularly the Diffusers-format) so as to:
1. Enable/expedite the creation of Huggingface spaces running on this model (I have a bunch of LoRA-testing Spaces, and I sometimes run one of them over this model).
2. Make it easier to train directly over this model using certain trainers. Namely, the ai-toolkit/Ostris trainer, via a Colab Pro notebook, which is what I've most often used thus far.
I've also made a few checkpoint versions of my WIP Historic Color model, via training over PixelWave Schnell v1 and then merging-in the LoRAs. I haven't yet posted any of these to Civit, but will in time, once I'm more fully confident in a version.
But tentatively, if anyone gets interested, here's a link to a fp8 safetensors of my current (third or fourth, but called V2) version of my WIP model:
https://huggingface.co/AlekseyCalvin/HistoricColorSoonrFluxV2/tree/main
@humblemikey regarding PixelWave for Schnell, you never mentioned a sampler setting. I suggest LCM/beta as it's x2 as fast as dpmpp_2_ancestral/beta for me and looks the same at 4 steps. Great model you have here!
and while I have your attention, how about doing a dev / schnell merge like FastFlux by nakif0968. It might allow PixelWave Schnell to have better skin textures.
Yess you've discovered the secret of VALHALLA! Try out my new workflow to see peak performance of this model!
@ericreator Aw man! I ran a grid in SwarmUI of every Sampler over every Scheduler on PixelWave Schnell @ 4 steps to find it out. All this time I could have just found your workflow and save 5ish hrs :(
@afterate2 we're on the frontlines of cutting edge research! We must sink countless hours into a minor improvements!
Amazing, thank you for sharing
v2 is my favorite dev model now now that i figured out the right workflow, but i want some more Schnell love, any chance for another round with Schnell for you?