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 an UNET only NF4 version of the full Acorn Is Spinning Flux V1.1 checkpoint for use with full steps. Though in my testing so far, it also puts out very nice images at 8 steps, though not as quickly as the actual Hyper 8 Step version.
FAQ
Comments (53)
i have to know, what's with the weird name?
It's an homage to an old favorite band from back in the 80s. Ace Frehley had a band, Frehley's Comet after he left Kiss. On their second album there was a song called "The Acorn is Spinning" which was an instrumental with a kind of background cinematic vocal track of a story about a boxer up and coming, rising through the ranks and questions on whether a fix was in on his big fight. It includes the boxing announcer calling the fight where at one point he takes a big punch and "the Acorn is Spinning." Just a cool song and when I was thinking about a name for the model, it popped in my head as a cool name for it.
@Seeker70 thanks for the explanation!
This great model is available online here.
https://app.giz.ai/assistant?mode=image-generation&baseModel=flux1&checkpoint=acornIsSpinningFLUX_schnellFP8V10.safetensors
thanks for sharing! anyone knows how i can achieve a fast model like the Hyper 8-Step here but without all the bias loras added in this model?
Sure, you can merge the regular Dev1-fp8 model with the Hyper Flux Dev1 8 Step Lora. That would then just be a faster version of the base model with all the normal censorship still intact.
Hey man! Love your work. This is my first attempt to use FLUX. I have downloaded the FLUX model Acorn is spinning v11hyper8, but I get errors and the model won't even load. I've been running XL and SD 1.5 for a long time without issue in SD FORGE. I'm far from an advanced user so I may have done something wrong. The Ckpt is in the directory I use for other checkpoints. What else is needed? Thanks
One, be sure to update to the newest version of Forge. If you are on an older install, you won't have all of the fields you need. Then you need to make sure you have the Flux VAE, a text encoder and Clip. The updated Forge will have a large drop down menu in the upper center where you actually select all 3 items in the same box. https://imgur.com/eI4j1ps
It should look like this.
This link will take you to get the VAE which will be calle ae.safetensors: https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main
This takes you to the text encoder: https://civitai.com/models/704402/flux-textencoder-t5-xxl-fp8-e4m3fn
This is the improved clip: https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors
The va.safetensors goes in the VAE folder. The T5 and the ViT-L-14 both go in the text encoder folder.
Also, the screenshot I showed you has the clip_l in it, but the ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors is the updated one that replaces clip_l. So you would instead have this one, plus the va.safetensors and the t5xxl-fp8 one all in that dropdown.
@Seeker70 This worked flawlessly. Thank you so much! Now to play with prompting in Flux!
@boejiden069 awesome. glad you got it up an running. :)
@Seeker70 Im using Forge Web UI, and this worked perfectly. U should consider teaching (;
@AndSudo02 Awesome, glad it's working for you. :)
Any chance of Acorn is Boning version?
Eventually there likely will be, but so far there just isn't sufficient resources to make one that would be of acceptable quality.
I can't get the NF4 version to work. please give proper instructions on how to install everything it needs to make it work. I'm using flux in ComfyUI
The easiest way is probably to download an NF4 workflow and then go into the manager to install missing nodes. The main thing is NF4 uses a different loader. here is on example: https://civitai.com/models/748920/simple-workflow-nf4
If you load the workflow and have any red come up, you just go to the manager and click "install missing custom nodes."
@Seeker70 I keep getting CUDA errors if I try to use NF4...
@ikwerner1432 Unfortunately I can't help much in that area. I rarely use Comfy. In Forge there is a drop down menu for Diffusion in low bits where you select NF4. Not sure in Comfy if there is a place for that or if it is just kind of built into the NF4 loader.
@Seeker70 I gues that is what the --low_vram setting is for, but it does not seem to work with NF4.
Can the DEVfp8V1.1GGUF fun in Forge? It throws errors for me. The non-gguf run fine for me
"RuntimeError: mat1 and mat2 shapes cannot be multiplied (1024x64 and 256x768)"
It won't work in forge unfortunately, only Comfy it seems :(
@Randmeist bummer...hopefully it gets fixed. I just started using Q variants and they're great
Unfortunately it only works in Comfy. I am not an expert on converting models to GGUF. Another user on Citivai converted them for me and I don't know why they don't work in Forge. I have had no luck when trying to convert them myself at all, so for now they just work in Comfy.
@BobSchmo Those variants are great but compiling times.... It's easier to use fp8, speed is the same, but no compilation
Nice! Will we be getting a Hyper Q4GGUF version 1.1 by any chance?
I can't say for now. I have been unable to figure out how to do the conversions myself. The attempts I have made don't work at all. Unless someone who is more of an expert than me comes along and does the conversion, there probably won't be one.
@Seeker70 I converted GUFF for you, please download and try it out. https://huggingface.co/sherlockbt/acorn-is-spinning-flux-guff There are two versions: Q5_K_M and Q4_K_M
@SherlockXinXin Much thanks. I am downloading and will test them out.
I was going to rant about all those aes, clip_Ls and t5xxls I had to download, but the quality of images is outstanding, I'm looking forward to work with the model.
Thanks. I hope you enjoy it. Most comments I get are people wanting the smaller download sizes and since all the Flux models use the VAE, Clip and Text encoders, most users tend to already have them installed.
[update]: I've updated the Q4 and Q5 GUFFs to v11.
I converted Hpyer Q4 and Q 5 GUFF for you, please download and try it out. https://huggingface.co/sherlockbt/acorn-is-spinning-flux-guff There are two versions: Q5_K_M and Q4_K_M
Why not convert normal fp8?
@Randmeist Q4 is smaller and it works better in ComfyUI, in Windows 10 ( it uses less RAM )
@jr81 I know that, I mean hyper is worse in details than normal 20 steps fp8 or any GGUF
@Randmeist I believe for most people the faster results are more important than the ultimate quality. For myself, I use the 8 Step 95 percent of the time as the difference in quality is very small.
@Seeker70 The 8 step is amazing on my 2060 Super 8GB. 43 second generation times, images still significantly more detailed than SDXL. Best model I've used for Flux so far.
@Randmeist I understand your point. I have now created non-Hpyer GUFF versions for Q4 and Q5 and uploaded them to Hugging Face. You are welcome to download and try them out.
@Seeker70 I've updated the Q4 and Q5 GUFFs to v11.
@SherlockXinXin OK, thanks. I will problem get them posted here too pretty soon. Been out of town visiting family. My page is getting pretty crowded. :)
@st3ph3nstrang3760 I have same card. What exactly of GUFF you downloaded and were i must put it?
@mihain The 8 step model I use is this one, https://civitai.com/models/673188?modelVersionId=862095 I tried the GGUF linked in the OP, the generation times were 2-3 seconds slower and I didn't really notice a quality difference so I just stuck with the original
@st3ph3nstrang3760 The main advantage of GGUF is smaller size / less VRAM usage.
If the whole model can fit in VRAM, it will be faster. If not, it may be slower than FP8.
In Windows, another advantage of GGUF is less RAM usage while loading ( at least in ComfyUI ).
A GGUF conversion will always lose a little quality in relation to the original.
In terms of image fidelity, Q8_0 will be better than FP8, if both are created from the same FP16 original.
But, in this case, we have Q4 and Q5, created from the FP8 original, so they may lose a little quality.
Hey @Seeker70 , if you haven't already, you absolutely have to try the recent Dedistilled models I've reposted here : https://civitai.com/models/843551/fluxdev-dedistilled
They allow cutting-edge prompt adherence even with many loras, negative prompt, and better quality through CFG setting. I swear I'm done with distilled models now that I have seen the light. I hope you do too, and that you make a version of Acorn on it ;-)
Thanks, I will take a look at it.
@Seeker70 Something about your checkpoints/models has hit the secret sauce. Dedisstilled needs and All in One package for Forge and Comfy. It would CEMENT itself the same way Acorn has.
@323f802 I am looking at it. I have made a test version using the Dedistilled FP8 model. So far in testing it is still pretty slow, but I am getting nice results at 28 steps instead of 60 so it is faster than using the dedistilled model on it's own.
@Seeker70 that's good new, what CFG are you using ? Someone has done a benchmark showing it's possible to decrease steps by using less CFG.
@323f802 I can't wait to get rid of distilled now. It is vastly inferior except for speed. Dedistilled can do everything distilled can do but better, and can do stuff that is downright impossible for distilled. It makes so-so LoRas brilliant, and now I started using it to retrain my Loras. It perfected a face lora I thought was really brilliant already, that I doubted could improve. The front faces were better, but the side views which were meh on the original turned to be excellent. This stuff is gold.
@valentinkognito365 Haven't experimented much yet, but so far most CFG 3.5 or 3.
@Seeker70 my favorite is CFG 8 and 60 steps
crazy results even with only 4 steps (dpm_adaptive sampler and karras/exponetial scheduler / max/base shift 0.20
dpm_adaptive doesn't use steps, it just plays with the sigma until convergence, so you might be doing 28 steps and never know it. I don't even know if it's uses the scheduler...not sure on that one.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.