Please consider buying my two albums Sisters of the Wicked and Danger Ahead, my book Dracula and also please check out my YouTube channel. Civitai won't support creators, so any purchase is a real help. Thanks.
DRACULA https://www.amazon.com/dp/B0FM4DXYPL
Sisters of the Wicked https://amazon.com/music/player/albums/B0DSJY3ZHY?marketplaceId=ATVPDKIKX0DER&musicTerritory=US&ref=dm_sh_CdyeuxIBNBDOjZ1hpB1PPJxT0
Danger Ahead https://www.amazon.com/music/player/albums/B0DB2DV354
Music In The Machine https://www.youtube.com/@musicinthemachine/videos
Lora Creation: https://www.upwork.com/services/product/design-flux-lora-for-creation-of-ai-art-1909718528813429477?ref=project_share
For the De-Distilled Model, Distilled CFG Scale should be set to 0. Then use actual CFG. For me 3.5 to 8 seem good.
If you like the models, please follow. The page is getting crowded as I keep releasing my model in as many formats and options as I can.
V1.1 Hyper 8-Step is now available and works in ForgeUI as well as Comfy. Schnell V1 Much faster, getting good results with 1-4 steps. Quality is not quite on the level of the Dev model, but much faster. So far it seems text is more consistently good though. New GGUF version of the Schnell model up now. Thanks to pkmngotrnr for converting them for me. Much appreciated. So far works in Comfy only at least for me. DEVfp8V1.1GGUF is now up as well. V1 version has European Flux mixed in at what seemed like a pretty low percentage, but it seems like it is a bit too strong. If you like this look, feel free to use this one.
This is my initial FLUX model. It is base on the flux1-dev-fp8 model. This is a merge only in the sense that I merged in some Loras that I like to make it more realistic. As with all of my models, the goal is the best photo realism that I can achieve. If it will do other things well, that is a bonus. Until I can learn a lot more about FLUX, and see if at some point true merges are possible, this is a pretty good start. I don't know about compatibility with all software. I use ForgeUI and I am able to run it fine. Most gens take about a minute and a half for me at 1024 X 1328. I use Euler with the Beta scheduler or the Simple. I haven't experimented much yet with other samplers. I typically use 24 to 28 steps.
Description
This is the full step version of Acorn Is Spinning Flux V1.5. I mostly generate at 1024 X 1280. For the full version I typically use 28 steps. I use DPM2 with Beta or Normal most often. I do think a few other samplers work but not as consistently well. I set distilled cfg to 0 (although it also works with distilled cfg from 2.5 or higher) and CFG to 1. This version has my Acorn Improved Nipples lora merged in. It doesn't completely eliminate all bad Flux nipples, but it cuts them down to a large degree. When generating non-nude images, if you get nipples showing through, it is helpful to use my Acorn Improved Nipples with a negative weighting. <lora:nipples:-0.99>
FAQ
Comments (44)
Absolutely smashed it with 1.5! Works well with my character loras and the body detail is quality! Good effort mate. Thanks
thanks a lot
Im just a little curious about the CFG settings in your preview images, you seem to have it at 0 or 1, and you have negative prompts on some of them?
How does that work, I assumed we never done CFG on flux, and just left the distilled CFG thing at 3.5 (or there abouts)
thanks
I took some of my old SDXL images using the png info tab to send it to txt2img. It automatically transfers the negative prompts as well. I don't think the negative prompt is actually being used in the generations though since the CFG is set to one, but when it outputs the final image, it is still transferring that meta data from the original image.
I normally have the Distilled CFG set to 0. Sometimes I set it to 2.5-3.5 but I don't notice it does much. However if Distilled CFG is set a 0.5 to 2 it seems to have very negative effects. For the regular CFG, if you want to use a negative prompt, you can set the CFG to 1.5 and it activates the negative prompt (at least in Forge). Then I think it actually does use the negative. However if you start setting CFG higher than 1.5 it starts to really mess things up.
Do you train your models with such prompts? May I ask why? And why not natural language? Doesn't natural work way better with Flux? Everything I've read says that prompts like (xyz:1.3) aren't good for Flux.
I don't train the models. They are made by merging Loras into it. I read that gone were supposed to be the days of weighting such as (xyz:1.3), but I have found it still seems to work with Flux at least for some things when there is not really any other way to give something a higher priority. I am not sure what particular prompts you were referring to, but on the latest model I was just running some of the prompts I used with SDXL to see how they work in Flux and they come out looking good, so I would say even though Flux can be prompted with plain English, it still seems to understand some of the old ways just fine as well.
@Seeker70 alright, alright. That explains a lot.
@Seeker70 Hey, what do you use to merge loras into checkpoint?
@zthrx I use Comfy UI. If you look on my profile I have the simple workflow posted.
With your latest version (AIS Flux V1.5), how can I train a lora using your checkpoint? I'm still exploring but I assume I'll need a full model with diffusers. Are there plans to make it available or can I do it myself from the pruned version?
I have never trained a Lora locally. I am not sure if you can use just the unet or if it needs the clip, text encoder and vae built in. I have no plans to upload an all in one model, but if you use Forge UI you can simply load the model, go to the Checkpoint Merger tab, type the model name and click Save Checkpoint. It will save a version of the model with whatever clip, text encoder and vae you have loaded, so it would be the all in one model in a couple of minutes. I just tested it on my computer. The resulting full model is a bit over 16GB, so if your machine has the horsepower to handle training locally with it, you can save it that way pretty easily.
I assume you can do a similar thing in Comfy as well, but I am not familiar with the exact process in Comfy nor do I have a workflow for it. Probably easy there too though.
Thanks @Seeker70! I ended up using this method: https://huggingface.co/spaces/John6666/sdxl-to-diffusers-v3/discussions/3
I don't use Forge but I use ComfyUI. Interesting to know that could've been done through there too, I might consider next time.
Cheers for the input + model! I'm really liking it
If you are using Flux gym this can be done very easily, flux gym offers you 3 models for training as default and there is a extra models file inside of the directory which you can edit and extra models. You can have it download it during the training process or store the model right on your device and add the correct path in the .yaml file. Training loras on finetunes work way better than training it on the basic flux dev model. If perhaps you find it confusing to change the path or modify the extramodels.yaml file in the flux gym, you can ask chatGPT for assistance, that's exactly what I did.
AIS Flux 1.5 not working for me, anatomy is all screwed up, usually just legs and arms in a ball. Not sure what I'm doing wrong. fluxdev1-fp16 works great for me, as do other flux checkpoints. I tried switching to the fp8_t5xxl but that dd not help.
Settings: Euler Simple/Beta. Steps: 24. CFG: 1-3. Running on a RTX 4090.
Update: It works ok when I use some of the detailed sample prompts here and adjust them to my liking. I guess it just doesn't do well with very simple prompts ("laying in bed naked").
My guess is the CFG: 1-3 might be the problem. CFG should always be 1 on Flux. Maybe 1.5 if you want to try negative prompts. If you go higher it starts to really mess things up. (Difference too between CFG and Distilled cfg. The distilled cfg for me is always either set at 0 or is set at 2.5 to 3.5. The regular CFG is always at 1 unless I want to use a negative prompt. Then would be set only at 1.5.) Also, for me at least, DPM2 Beta/Normal work much better than Euler.
hi, i noticed the latest version has a high tendency towards nudity, sometimes i make a promp for someone with specific clothes and the image comes with said clothes but open showing the body or with nipslips, this is not a complain just something i noticed
Yes, it definitely does that sometimes because it was created using my Acorn Improved Nipples lora to help correct the awful Flux nipple problem. In part it was created for those who want better nude images from Flux. I have found though that if you want images specifically without nudity, you can use that same lora at a negative weight, such as <lora:nipples:-1>, and it will almost always prevent unwanted nipples from showing up.
Absolutely stellar work, man! Congratulations on the quality! I don't even get banding on high res images. So good!
Thanks a lot.
Why did you delete the De-distilled version? Overall this was one of the best versions of Flux on the site...
I am considering finding a new home for my models since CitivAI has abandoned supporting creators. They have given up any pretense of having a creator program. As creators, we don't even get the tips people intend to give us. We get the buzz but have no mechanism to actually cash out the buzz so it's just useless. I am still considering what to do. I give options for people to help support the work, but no one ever does as long at it's available for free.
@Seeker70 That's sad to hear. I'd like to continue following you and your works. Will you tell where to move to when you decide on a new platform?
@Akalabeth Sure. I am not sure if I will move them. I was looking into TensorArt but in their program, only super tame checkpoints are eligible for their paid program. It may be that there just isn't any viable place for creators out there.
@Seeker70 Maybe we should create our own website? Perhaps something federated. It might not be easy, but it would probably be better for the whole community long term. We already have Pixelfed for sharing pictures, maybe it could be modified to add the functionality we need. Anyway, thank you for your amazing work!
@Seeker70 To get money, you should do little more than silly merge checkpoints in WEB UI which can be done by anyone. Same story like with SD1.5 and countless mixes, but in reality there only 5-7 really quality trained checkpoints, others are just mixes of mixes remixed with stupid loras added.
@littlefluffyball I am only concerned with Flux at this time. As far as I have seen, there are actually zero good trained Flux checkpoints. That is why I do merges. Right now, merges work better than training. I am involved with a group working on training a better Flux checkpoint, but it takes a lot of time and work and people want models to use in the meantime.
@Seeker70 Pixelwave is the only well-trained Flux checkpoint on Civitai. It is great for creating artistic images. For realistic images, Flux-dev with LoRas or your merges is the best choice.
I keep getting the same face with every generation is there any fix please
Flux has a fairly standard face a lot of the time. Make sure you are not using the same seed. I don't seem to have a problem with getting the same face. Of course you can use Loras to get different faces also.
@Seeker70 it is randomized , but faces are too similar the model really needs more diversity, used other flux models it wasn't as noticable as this
@CLAAN3008 Not sure what to say. I don't get that problem when I use it. The base of the model is the Flux 1D fp8 model. No face Loras were added, so any face qualities Flux has should still be in there. The only thing I can think of would be if you were using my very first Acorn Is Spinning Flux V1.0 which had a European Faces Lora mixed in which was too strong, so V1.1 and all since do not include that.
Does anyone else have Problems using this Checkpoints with a Character Lora? Maybe I am missing something in the settings but it seems like no matter what CFG or Strength of Lora I put in, it never really comes through and I get very similar faces that have nothing to do with my Lora. With the normal Flux D FP8 the Lora works great
Which version of the model are you using? I have several character Loras that I use and all work perfectly with all versions of the model. CFG should always be set at 1. You can do 1.5 if you want to use a negative prompt, but anything over that will really start to mess up the output. Also, Distilled cfg I usually set at 0, but works pretty good from 2.5 to 3.5 as well.
Im using the FLUX pruned FP8 Model. Could you maybe Drop an example Workflow with all these settings? Im still pretty new to generative AI. Thanks Brother!
try with Diffusion in Low Bits: (automatic fp16 lora) in forge good luck.sorry my english ist bad
@mclovinofm417 All of the models are pruned fp8 models. They are either V1.1 or V1.5 and some are Hyper 8 step models, while others are full step models where you need to use 20 steps or more. Others are GGUF or NF4. All of them have different ideal settings, so it is hard to say without the more specific model name. Also, unless you use ForgeUI for your generations, I won't be able to give a lot details as to the setting. I don't use Comfy, so I don't have a saved workflow.
@Seeker70 Oh sorry I wasnt to specifiv Using the V1.5 Pruned FP8 Model that has 11.08 GB. Been using it in ComfyUI so far. So if you maybe have a JSON with a Workflow I would be really thankfull bro!
@mclovinofm417 I don't use Comfy, so I don't have workflow for it. I pretty much only use Forge UI.
@Seeker70 Ah Damn, what settings would you recommend for this Model? Especially if I use Loras aswell. Also what Scheduler and Sampler work best?
@mclovinofm417 For me, I normally generate at 1024 X 1280, typically 20 steps if it's the regular model or 8 to 10 steps if it's the Hyper model. DPM2 sampler with Beta or Normal. I use Distilled CFG at 0 and regular CFG at 1. If you use the De-Distilled version of the model, that one is different. For that, Distilled CFG is 0 and regular CFG 3.5 or higher and seems to work best with Euler sampler with Beta and I use 28 steps or more on that one.
AIS is easily my favorite checkpoint. In terms of accuracy, not messing up (face) LoRAs, and built in NSFW, it's most versatile checkpoint out of the two dozen or so I've used. Keep up the great work!
Thanks. I am glad you are enjoying it. Many hours of work have been spent on this. It's nice that even though citivai has totally failed to compensate any but the very top creators, at least I can know that people are enjoying the model. Thanks again.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.