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
Schnell FP8 GGUF. For now I can't get it to work in Forge, but it does work perfectly in Comfy UI. Hopefully soon Forge will be able to run all the GGUFs. For now it runs some and not others.
Thanks to pkmngotrnr for converting them for me. Much appreciated.
FAQ
Comments (122)
thanks so much!!! seems great!!! I would say a GGUF of the dev fp8 is required! ;)
Yeah, I have it. I am running my test images and will upload it tomorrow. Too late to start it right now. :)
are you planning on releasing q6 or q8?
No immediate plans. I haven't figured out how to do the conversions yet, so for now I just have the Q4_0 ones that pkmngotrnr converted for me.
here is the q8 version: https://pixeldrain.com/u/52TBTf7A
Could you upload the Dev fp16? I can quant it to Q8_0, which IMO is the way to go!
I don't have a Dev fp16 version. The fp8 was the biggest one I am able to run on my laptop, so I created it from the Dev fp8 to start.
here is the q8 version: https://pixeldrain.com/u/52TBTf7A
@pkmngotrnr I appreciate you uploading it, although at least for me the download speeds are insanely slow.
What do you use to quant? I've trained a bf16 flux dev, but I go OOM when I try to use it with loras, LOL.
until now GGUF doesnt support LORAS on WEBforge-UI ;)
At this time my GGUF models don't work in Forge UI, at least for me. They work well in Comfy UI and all of my sample images were made in Comfy. I have other GGUF models that do work in Forge that are of the same Q4_0 type, so I don't really know why they don't work. I'm not an expert in any way on GGUF. A few weeks ago I had never even heard of it. If anyone who is more of an expert on the subject would like to see if they can get them working in Forge, please do.
My regular fp8 versions of both Acorn Is Spinning Flux V1 and V1.1 as well as Acorn Is Spinning Flux Schnell V1 all work well with Forge and Comfy.
The problem is precisely in this model, all other guuf models work for me on forge.
Same error for dev and schnell.
Forge: "AssertionError: You do not have CLIP state dict!"
ComfyUI: "CLIPTextEncode
'NoneType' object has no attribute 'tokenize'"
Anybody?
https://imgur.com/eI4j1ps
For Forge you have to have the Clip, the VAE and the Text encoder all selected in the box at the top. If you don't have the box, you may need to update Forge. For Comfy there is also a place to select the stuff, but I don't have a screenshot of that because I rarely ever use Comfy.
For Forge, just use this guide:
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050
anyone with the comfyui solution? im having the same problem :(
@eXAiPic I would say the best option would be to look on the site here. There are a lot of good Comfy workflows that people have posted. If you load a workflow then go into the manager, you can automatically install any missing nodes that that workflow uses that you don't currently have installed.
@Seeker70 Just found one, its all working! Amazing results!
This model is my favorite for generating NSFW photos with FLUX thus far. I've created a workflow and wildcard setup for generating infinite realistic amateur photos here!
Thanks a lot. Your images look great.
Do those wildcards work with the Forge/A1111 wildcards extension?
Hi, I want to try the gguf dev model but I keep having the following error:
I am using the unet loader for GGUF files as well as the Dual clip loader for GGUF. Any idea how to fix it ?
Error occurred when executing UnetLoaderGGUF: GGUF magic invalid File "/workspace/ComfyUI/execution.py", line 152, in recursive_execute output_data, output_ui = get_output_data(obj, input_data_all) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/execution.py", line 82, in get_output_data return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/execution.py", line 75, in map_node_over_list results.append(getattr(obj, func)(**slice_dict(input_data_all, i))) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/custom_nodes/ComfyUI-GGUF/nodes.py", line 258, in load_unet sd = gguf_sd_loader(unet_path) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/custom_nodes/ComfyUI-GGUF/nodes.py", line 39, in gguf_sd_loader reader = gguf.GGUFReader(path) ^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/venv/lib/python3.11/site-packages/gguf/gguf_reader.py", line 95, in init raise ValueError('GGUF magic invalid')
DevFP8 V1.1 is working but the DEVfp8V11 GGUF and Schnellfp8 GGUF give me the following error: RuntimeError: mat1 and mat2 shapes cannot be multiplied (4032x64 and 256x768)
Any idea what the issue might be? I am using the latest version of Forge WebUI, 816x1024, ae.safetensors, clip_I.safetensors and t5xxl_fp8_e4m3fn.safetensors.
I just saw your post saying the GGUF models are not working in Forge but I will leave this up in case anyone has an answer.
I encountered a similar problem in the original version of the dev fp8 model because some nodes were not compatible with the flux model. After I replaced them with nodes that are compatible with the flux model, the problem was solved.
@march213 I assume the nodes you mention are for ComfyUI? I am using Forge webui.
@stormlocks368 Yes, Comfy has different nodes to load Flux and GGUF models. I am not sure on my GGUFs if the problem is with the particular conversion of the GGUF model or if Forge just doesn't work yet will all GGUF model types. I do have some Flux GGUF models like a basic Flux1-Dev-GGUF that does work in Forge, so again, not sure what the issue is.
@lexy007 I have used both of my GGUFs in Comfy and they work flawlessly, so it's not purely a problem with the models, but likely somehow a compatibility issue. Anyway, for me it has become somewhat irrelevant now that I made the 8 step and NF4 version because they work at least as well and faster than the GGUFs anyway. Still, I would like it if the GGUFs could be made to work in Forge since Forge is my preferred generator. I am not an expert on using the conversion scripts and someone else converted them for me, so I don't really know if there is some kind of settings that are done during the conversions that make them incompatible with Forge or something. If I get it figured out how to do them myself, I might try to redo them.
@Seeker70 thank you very much for the answer
a Hyper 10 steps Unet version would be really cool
I think you got one of the best flux checkpoints here. It would be great to have an hyperdev 8-steps quantized version (addictively faster). Surely the folks that released these checkpoints could help ?
Thanks. I have been especially loving my Dev(GGUF) version even though I can only use it in Comfy for now. Not sure about the Hyper. I've done a little experimenting with the Hyper 8 step one, but so far none of the ones I have tried would work, even in Comfy. I'm sure I will keep looking into it.
I've been playing with the GGUG hyperdev 8-step Q4 and Q5_1 and honestly I was skeptical at first, but I have to say it holds its promise. Most pics come out with very good quality in 8-steps. Can always make it 10 or 12 steps if required. I've gained like 40-50% time when generating, which is really a game changer. I can now generate 1152 * 782 pictures in less than 30s, sometimes even 20s. Before that it would take almost 1 minute for a similar result. I'm using Forge. I've actually just found out this is also a Lora so I will try Acorn + the 8-step Lora. Although, I don't know if using a checkpoint + this Lora is different from having the checkpoint merged with the Lora, both in terms of speed and resulting quality.
@valentinkognito365 I can't say definitively, but I expect it will be slower because at some point during the generation, the Lora has to be loaded and run. In my experience, using a Lora for me, adds at least 30 to 40 seconds to the time. Obviously some of that would be dependent on the system specs you have.
@Seeker70 well I have started trying with the optimized 8-step Lora and it DOES work with your checkpoint. I generate in 30 s with it as well. Quality is very ok at 8 steps, but 10 or 12 steps could be good as well to remove some grain. Are you using Forge ? There are settings to keep loras and models in VRAM. It doesn't always work (in the logs) though. Models, including loras are loaded in like 5 - 7 seconds, then it immediately starts the 8 steps. If I'm not changing the prompt or anything, further generations will take less time, 24 on average, sometimes even 20 s. ComfyUI was excruciatingly slow on my setup (and I have a powerful laptop with RTX4090), I had the same issue where I would wait ages just to load models and stuff. Since I switched to Forge, I never looked back to ComfyUI and its overcomplicated workflows for Flux
Also, I have 16 Gb VRAM but I prefer to use Q4 or Q5 quantz, using only like half of my VRAM, because the VRAM unloading is often not complete between generations. This way it's easier to keep it in VRAM or at least avoid swapping to CPU memory. I might lose some quality, but it's worth it for me to have more speed
@valentinkognito365 I use Forge if I am using my regular models. If I am using the GGUF models I use Comfy for now because my GGUF ones don't work in Forge, at least for me. So for now I am using both Forge and Comfy. I do prefer Forge when I can use it.
@valentinkognito365 I haven't seen that Lora. Do you have a link to it?
@Seeker70 I can use your GGUG in Forge but not with the acceleration Lora (might be no loras at all, would have to check)
@valentinkognito365 That's strange. Even after updating Forge just now. I still can't run it. I get: RuntimeError: mat1 and mat2 shapes cannot be multiplied (4032x64 and 256x768). Do you have to change any settings to run it in Forge.
@valentinkognito365 The hyper 8-step is up. Thanks for the suggestion.
@Seeker70 that's really good news, thank you so much. I average 25s for 8 steps with the hyper version compared to 30s average with the FP8 + Hyper lora (optimized version). This few seconds gain is very appreciable, especially on the long run. Quality seems the same. 10 steps now take the former 30s and is cleaner, if not in a hurry. Now it's definitely the best flux checkpoint available.
Btw, do you have any plans for Acorn is Boning - Flux ? just wondering :-)
@Seeker70 I think I was also getting the "RuntimeError: mat1 and mat2 shapes" error with the GGUG version + the hyper lora. It's fine now with the version you've just released.
Maybe a GGUG hyperdev version would be possible though, as hyperdev would be blended in the checkpoint like you did. That way perhaps there wouldn't be a need for Forge to "fit" / "match" the Lora with the checkpoint, resulting in this error ?
I'm fine with the latest version though, but others might be interested by the GGUG + Hyperdev combo (I would as well if you release it :-)
@valentinkognito365 I get the error with just the GGUF checkpoint alone. Neither of my GGUFs will work on my Forge at all. It's weird because I have a straight Dev1-FP8-Q4_0 and it works perfect in Forge. Anyway, now with this 8 Step version I don't expect I will use the GGUFs much anymore.
@valentinkognito365 Not sure yet on Acorn Is Boning. Most likely I will have to wait until there is quite a bit more Flux content out and a lot of better stuff. Most of the stuff out so far is pretty sub-par stuff.
@Seeker70 yes, 10 steps with the hyperdev version is neat. I agree about the flux content, too many cheaply made Loras out there... I'm considering making my own NSFW loras with headless source pictures and slim body types. Tired of the loras messing up with face and/or body type (ugly + somewhat chubby + big tits).
I don't know how you've done for your checkpoint, but thanks you did. Now that I've heard about this binary mask thing among other isolation techniques, I'm starting to think the issue is a lack of knowledge... but then flux is recent...
great model that actually works and with some minor tweaks it could the best iv tried sdxl and version before and never fail the test
Thanks.
The model I am currently using has the best overall experience. Nice work!
Thanks a lot.
In terms of quality and realism which is better v1.1 Hyper 8-steps or DevFP8 v1.1? I want to know specific purpose why you update this model. I have low net speed so want to make sure i download the best one.
I think for quality they are very similar. I update it because many people want to generate images in less time which the Hyper 8 step version does. For absolute quality they are very similar because with the Hyper 8 you can still set it to the higher steps if the lower step image isn't good enough. Though if you look at my sample images, the majority of them were done at 8 steps. I found the ones with text needed a couple more steps to consistently get the text right. For me the regular DevFPTv1.1 takes about 1.5 minutes per image and the 8 step takes about 35 seconds, so it is much faster and probably 90 to 95 percent of the quality of the main model at 20 steps.
@Seeker70 Thanks for replying before downloading v1 Hyper 8steps is complete. I struggle a lot with my rtx 4070, 32gb ram. Flux dev and schnell took me 2 - 3 mins even with 712x712 res. schnell is not fast honestly only 2-3 sec difference. So i download so many alternative checkpoints 120 gb. Check n delete many times n 80% of them looks like ai generated image and oily plastic body such as Flux schnell, Flux1-dev-bnb-nf4, NewRealityAllinOne (the worst model) etc. Last night i check yours n quite impress with quality and nsfw even without Lora. Today I tried this 1024x1280 res took me 56 sec. It looks a little oily plastic. So I use Lora and reduce by 30%. I like this it follows prompt well and + nsfw without Lora is a big bonus. Thanks a lot for your SERVICE 👍
@haidensd58757 Hi, I have the same PC specs as you and can get faster completion times with Flux dev. Do you use FORGE? I'm using that and overall it seems faster than comfyui. Multiple loras do slow it down however.
@jonog247634 No i didnt use forge. I have Stable swarm which is also really fast but sadly dont support Flux. I'll try Forge today. Thank you for suggestion 🙏
@haidensd58757 I would second the idea of using ForgeUI. I switched to Forge quite a while ago and now I will only use something else if I find something Forge won't do, like for some reason I can't use my GGUF models in Forge. With my new 8 Step Dev model though, I don't expect I will use the GGUF much anymore.
I am surprised by those times you get with Dev and Schnell. I use a 3080 with 16GB VRAM and 32GB RAM and with the full Dev1 model it takes me about 1 min, 45 seconds generating at 1024X1280 and 20 steps. With my new 8 Step here it's usually in the vicinity of 33 to 36 seconds at the same res with 8 steps. If I bump it up to 10 steps it goes up to around 45 seconds.
@Seeker70 I just installed Forge UI 20 mins ago following suggestion from jonog247634 Transfering all my Auto1111 datas to new Forge. Man i love it 🩷 feeling the same home as Auto111. Its fast really 🚀 Its a Game changer. I think am not going back. I checked your Flux AcornIsSpinningFLUX_v11Hyper8Step with vae ae.safetensors + clip_I + t5xxl_fp16.safetensor + Lora (XLabs Flux Realism LoRA). It took ⌚55 sec to generate single image of 612x868 res. 1024x1024 will take 1.5 minutes. Using comfyui setting DPM++2M SDE Exponential 8 Steps and 1 CGF didnt work. I have to stick with default settings Euler steps 20, 3.5 cfg.
😊Btw am happy with your new model v1.1 Hyper 8 steps it follows prompt very well + nsfw included. I think you can really do well for ForgeUI in your next update like fixing Skin Texture to be more natural+hands hard to tell Ai or real. ⌛ Forge UI is robust, fast, reliable, very user friendly, brings back our good old mother Auto1111.
@haidensd58757 Yes, Flux doesn't work well yet with most samplers. In SDXL my favorite was DPM++ 2M SDE Heun Exponential, but it doesn't work in flux. Euler is the main one for now. For me I use Euler Beta mostly, but Simple, SGM Uniform, DDIM and Normal seem to work as well. Probably some others, but for me, Euler Beta is the best for now.
@Seeker70 Agree gotta stick with Euler for now and hope Flux will bring more optimization for mid range GPU without compromising quality. Flux right now seems to struggle with realistic skin textures more focus on Text and Hands. But thanks to Loras. Thank you for hard working creators like you, making CivitAi a better world.
@Seeker70 Yeah I've been using Hyper Models for a bit and I find 12/6 (Base/High Res) with sub 0.4 denoising is a decent baseline in terms of coherence and quality.
How do I use it in comfy? I downloaded hyper 8-step, placed it under models/unet and used "Load Diffusion Model" node as the loader. Comfy just crashed on me... what am I doing wrong? Thanks!
It would be ideal if someone could post a comfy workflow that works.
https://civitai.com/images/30051408
Download this image and drag and drop it into your comfy. It will populate the workflow I have been using. It has 3 different loaders. The one hooked up works on the 8 Step. One is for GGUFs. The other is if you are using a flux NF4 model. Just hook up whichever you need. By default the 8 step is already connected.
@AskarK thank you. It seems to use a 3-rd party "Mixed Cast Flux Loader" node that I'm missing. Is there no other way to load it using the standard nodes?
@multikertwigo Yes, you should be able to double click on any blank area to bring up the search. Search for Load Diffusion Model and that loader will work also. Also if you use the Manager you can also bring it up and hit the "Install Missing Custom Nodes" button in the upper part of the center column and just install the one that is missing. Either should work.
@AskarK Tried that, it just crashes. I think Hyper NF4 is just broken.
@Mescalamba You could try re-downloading it. Maybe it got corrupted somehow during the download. I have used it quite a bit and it has been working perfectly.
@Mescalamba my problem was insufficient paging space. Both Comfy and Forge just crashed when I tried to load big ass models (as in, .safetensors :) ). Look at the messages in event viewer. If it's saying something along the lines "virtual memory resource is exhausted" (sorry, don't remember the exact wording) - that could be it. I enabled paging file on another drive and cleaned up some space on the first drive that used to be the only drive for the page file. It was running low on space (<20GB). Then the problem disappeared.
Yes thanks for the hyperFlux, a Unet version will be much better for my dd ;)
cool
The V1.1 version of Hyper 8-Step can be a great option. So far, I have tried it with the standard Simple, Normal schedule types - the result is a bit grainy. Has anyone tested in depth what combination of sampler/schedule type gives the best result?
For me the best is Euler Beta. DDIM, SGM Uniform also work.
Are you still using Euler and Beta and the same 1024x1328 resolution for the 8 step version? I'm getting similar generation times to you, but the skin and other details seem to be much more smooth or low quality then the images you posted.
Lately I am using 1024X1280 mostly. The samples mostly use Euler Beta at 8 steps, but a few use 10 steps. I found ones with text sometimes needed a couple more steps to be more consistently right. Done in Forge UI. Some might be at 1024x1328 if I used PNG Info to transfer the prompt from a previous image to txt2img and didn't notice it was at that resolution. 1024x1328 still seems to work fine, but I had went to that res specifically in SDXL because at least for me, that seemed to improve hands. In Flux, hands are normally much better anyway and I prefer the images not to be quite as tall.
The model is great, as i already tried q4 and fp8 hyper version, but it's almost impossible to generate other ethnicity, sometimes i got luck getting a bit brown skin or asian face but most of the faces are very european, and has the same face, i thought it's only on your v1 euroface model.
edit: after messing around with cfg and guidance, now i got what i want, thing is i misinterpret flux guidance as cfg scale on comfyui.
@N3kor0 May I Ask what settings worked for you? I seem to run into the same problem...
@BKFEEST for me i set cfg to 15, and using dynamic thresholding set to 1, for guidance i set it to 1.5-5 max, with heun sampler, 28 step, but i think flux is quite different because it uses natural language as prompt, so adding western trait like blonde will (most likely) give you european/western person, idk why, but it works with other hair colors, i'm still trying to learn how to prompt as well, because flux is total opposite of pony xl in terms of prompting.
I have had good luck with all different ethnicities, but you usually do have to be more descriptive and use plain English for prompting. For different ethnicities, I tend to use specific countries a lot at least when it's a country that has a high percentage of one ethnicity. I might use Ethiopian or Sudanese if I want a darker skin or if I want Asian I usually specify Vietnamese or Korean or whatever. Biggest thing overall though is getting rid of the idea of tags. Just say what you want and it works the best.
I rarely use Comfy, but in Forge, I almost always leave the CFG at 1 and usually for most things I set the guidance at 2.5.
@Seeker70 yeah, after a long time pony user, i'm still adjusting how to prompt in flux as they're total opposite....
I am not getting any luck with this model with character loras. It changes the face too.
Your loras are the problem. I trained 9 flux.d faces Loras myself recently, I've tested them ALL using Acorn, I can tell you it does NOT modify the face. At all. I had incredible and consistent results with my Loras. Unfortunately, most civitai flux.d Loras are quickly made and just bad.
A few weeks ago I would read that it was a bad idea to use more than 20 - 30 source pictures to make a character lora. I just released a character lora made from 300 source pictures that works very, very well, putting to shame the Loras I had downloaded on civitai so far. Unfortunately, everyone keeps doing that. And it might get worse with that 5-minute express training online.
What to expect from some Loras being trained on 10 pictures, on in 5 minutes with only 1 epoch ?
All NFSW loras I've tried also suck because they modify faces, faces being not excluded from the training.
I've found out flux is an impressively good learner, but we need to tame it.
I'm sorry to say that, but consider training flux loras yourself if you want good stuff, it's worth it.
I'll second the LORA's definitely work sentiment. Obviously the LORA is working off of the data from the model that would generate a specific face, but all it takes is proper weighting and it comes through great. I find I have to weight some LORAs even less with acorn than with dev. I also use as many images as I can for my LORAs as well, civitai will accept up to 1000 for a job.
i agree, i feel like the only way to get good results of character models in models that dont work well with them like this one does is to overtrain the character lora. for me the only other model ive somewhat liked with my loras was flux unchained. I also feel like it has to do with loras being merged into it instead of being finetuned.
@plus1 I did not like other nsfw model or lora precisely because they were either face altering or learning towards big boobs which is not my taste. Acorn is the most balanced and finest one I've used so far. The only draw back I found is it is so-so for pussies, but using a pussy lora fixs that easily. Most chkpt are merged indeed. Not sure about it abeing an issue compared to trained thought. First thing to make sure of is your lora is actually STRONG by itself indeed, and many civitai ones are weak. Always check step number and epochs on the lora page. 8 - 10 epochs / something around 100 steps per source pics are signs of a strong lora. Dont even bother if the face is not true to the character.
I have a project of merging my everyday loras locally with chkpt i use such as acorn, probably would help, at the very least for performance
@valentinkognito365 i ill give it another try maybe i just need to lower the lora strength or something
@valentinkognito365 oh yes and i just assumed its harder to balance face with lora than training just because you have to merge loras at a certain strength and i just thought eventually it would probably mess somethings up i may be wrong though
Awesome. Appreciate the NF4 Hyper. Will give it a workout over the weekend.
I got blue image, any ideas? with NF4 from ilisvyael it worked
Probably need more information. What software are you using? Do you have VAE, text encoder and clip all selected? For instance, I use Forge UI, select VAE, text encoder and clip, choose BNB NF4 in the low bits drop-down. Also make sure you use a Sampler and scheduler that work. I use Euler Beta mostly.
I had the same problem with any FLUX model before. As I understood, it's connected with lack of swap memory (you need more than 40k MB free space in every logical disk you have if you didn't specified the exact disk for swapping).
The problem described somewhere in ilisvyael's Github page, unfortunately I've forgotten the exact link.
@Seeker70 Thanks! It would be great if you could list the exact VAE, text encoder and clip that you are using!. I also use ForgeUI
@mphobbit Yes, this is under control but thanks anyways!
@AiLEarner this is my issue too I think.
I would love to see your checkpoints converted to Q6_K format. That format is darn near as good as full Flux Dev, but saves tons of storage. Great images.
The quality of my 8 Step NF4 is pretty amazing too and I think it's as least as small as the Q4_0 version. And it works in Forge as well as Comfy. Unfortunately I have had no luck with doing the Q formatted GGUFs myself. None of my attempt at those worked.
I love acornIsSpinningFLUX_devfp8V11 and acornIsSpinningFLUX_devfp8V11_Q4_0.GGUF. I tried hyper, but found that the generated images were closer to standard dev FP8 than your acornisspinningFLux_ devfp8v11.
If making a Lora locally, do you think it would be more successful if it was trained with your model as the base instead of standard dev FP8 or Q8_0? Any recommendations on making LORAs that will play nice with the model and not override the qualities of the base model?
I can't help much in that area. The only Loras I have made were done on the Citivai trainer with the main version of Flux. They seem to work well for me, but I really very rarely ever use Loras. I have never trained a Lora locally. Others have told me that they have used non-base models to train Loras, but I have never tried that myself.
Lately the one I use the most is the Hyper 8 Step version. In my experience the output is basically on par with the regular with my acornIsSpinningFLUX_devfp8V11 but in half the time or less.
@Seeker70 For whatever reason every time I have tried using the 8 step version I ended up with images that were less photorealistic. They were good, but when I would match up my seed and all other settings and change between, AISdevfp8v11, devfp8, and AISdefp8v11hyper, the results of the hyper version would be much closer to that of the base devfp8. For hyper I tried with my steps down around 12-24 to see what the difference would be and for whatever reason it always seemed closer to the base model than your AISdevfp8v11. Maybe my issue is I never actually tried only 8 steps, because I think you said for good text you would often need to do 10 or more for improved performance. Maybe on higher steps it gets more like base than AISdevfp8v11? In either case for doing 512x1024 images with controlnet to look for ones I like for upscaling it only takes me like 12.5 seconds per generation for AISdevfp8v11.
@SuoixDs I don't like to compare things with the same seed. A seed that works on one model doesn't necessarily work with a different model. Even though all of my models are Acorn Is Spinning and start with the same base, anytime you add something to it, it is really a different model. I will keep a prompt for images between models but I always prefer to randomize the seed to get a better image specific to the particular model. If you try to keep the same seeds the images tend to get worse and worse over time because the seed was one from like 4 model iterations ago.
Of course everyone has whatever workflow they prefer, but for me I find the results are always better if you start at higher resolution even though it's slower, than to do low res then upscale. For one, I have found that most models just don't produce as good of quality at low res. If you upscale something of lower quality, it kind of magnifies the poor quality in my experience. Flux was trained at 1024X1024 so for me I always start with one of my dimensions at least 1024. I actually get some amazing result generating at 1920X1080 as well when making images for my AI music videos on YouTube. I just love Flux. :)
I'm afraid I cant get any of these working in Forge, it just times out and disconnects.
I have t5xxl_fp16.safetensors, is this the issue?
That one works. You also have to have selected the ae.safetensors and the clip_l. They are all selected in the same drop down menu as long as you have them all downloaded and in the right folders. The t5xxl and the clip both go in the models/text_encoder folder.
it may disconnect (crash) if there's not enough vram or ram to do computation.
Works for me out of the box in forge. But cranking it up from 8 steps to 16 makes a huge difference. BTW why is my DEIS + DIMM setup n't working anymore? Find Euler a bit boring.
Can you also make a GGUF version of the 8 step model ? Thanks !
GGUF Dev Q4.0 not working at all in forge - RuntimeError: mat1 and mat2 shapes cannot be multiplied (4032x64 and 256x768)
mat1 and mat2 shapes cannot be multiplied (4032x64 and 256x768)
Yes, I stated this in the version notes. The GGUFs only work in Comfy as far as I know. I can't say for sure about other generators such as Invoke, Swarm and others I don't use. I primarily use Forge and I use Comfy if I have to for a few things such as these GGUFs.
@Seeker70 This particular model GGUF is not working, but original flux dev converted to GGUF variants are working... It's sad that you don't have nf4 for 20 steps like bnb-nf4-v2...
@RoninRex I probably could do an nf4 version of the regular model. I think that is one I can do on my PC. I will look into it.
@RoninRex I do wish my GGUFs worked in Forge. I don't know how to convert those. I have tried but can't get the tools to work for me. Another using here converted them for me. They stated at the time that Forge doesn't work with some of the versions, but I am not sure exactly what the problem is. I also have the Dev1.fp8.Q4_0.GGUF that works in Forge, so maybe it is some issue during conversion. I wish I knew.
@RoninRex The full version NF4 is up now.
@Seeker70 Thank you!
@Seeker70 I assume you didn't include ae,clip and t5-xxl(fp8) into nf4? Original flux1-dev-bnb-nf4 size is 11gb and includes those
@Randmeist No, UNET only. I pretty much assume by now that people already have the other things installed.