V4
Alright, so what’s new in this version? I cranked up the aesthetic dial, added more diversity in ages, and improved how it handles Asian features. But - because there’s always a but - I did notice the hands got a little wonkier. Eh, can’t win ‘em all.
I highly recommend pairing this with my LoRAs, like the realism amplifier, 2000s analog core, and others, since this checkpoint works best as a base for stylized LoRAs. Might do one more version (because, let’s be real, I kinda scuffed v3 and v4 a bit), but first, I’m diving into fine-tuning Flex.Alpha.
This time available versions: bf16, fp8, q8_0 - pruned fp16 name and q4_k_m - pruned fp8 name
P.S: Don't use my UltraRealPhoto LoRA with this checkpoint - it has a huge impact on style, so image become overbaked. If you're using the UltraReal Fine-Tune, go with Realism Amplifier instead for the best results. UltraRealPhoto LoRa was created to fix crappy shadows, light and faces, but all that stuff already baked inside checkpoint, you can just add amplifier for better realism
V3 Update (Experimental)
This release marks a step forward, although it’s still very much a work in progress. I focused on improving several key aspects, such as nudes, feet, and lower body anatomy. While the results are better than before, they’re not yet at the level I’m aiming for. That said, this version brings noticeable quality and texture enhancements, offering more detailed and refined outputs compared to the previous versions.
Recommended Settings:
CFG Scale: 3 (instead of 2.5 used in earlier versions)
Steps: 50 (helps with stability, though some minor instability remains in hands and fingers)
CFG 0.9 vs. 1.0: Lower CFG on 0.1 or even 0.2 may sometimes improve some details (may not improve, so feel free to expriment with this too), though it might take longer to generate.
Regarding nudes: they are still not working as intended, but I’m actively working on this issue and expect to address it in the next version.
The good news is that I already have the datasets prepared for V3.5, which I aim to release much faster than the gap between V2 and V3. With more experience and feedback from this version, I’m confident the next update will deliver significant improvements.
As always, I truly appreciate your support and feedback - it’s invaluable as I continue refining this project ❤️
P.S.: I feel like the more I fine-tune Flux, the more it degrades in other areas. Also i thought about trying finetune Flex Alpha (project looks very promising)
What's New in v2.0?
Enhanced Anatomy: Hands, feet, and poses have seen major improvements, offering more natural and accurate results. Say goodbye to overly distorted limbs!
Improved Textures & Quality: Upgraded skin details, richer textures, and sharper results overall. Blurred images still happen occasionally, but much less frequently than in the previous version or when using LoRAs alone.
Improved Text Rendering: Efforts have been made to improve the generation of text in images, and it’s much better than before. However, artifacts can still occur, and strange symbols might sometimes appear instead of readable words. This remains a work in progress.
Expanded Dataset: A larger and more diverse dataset (1800 images) introduces better balance across styles, lighting, and compositions.
Added Checkpoint Variations
To ensure compatibility with different workflows, I’ve included multiple checkpoint variations:
BF16
FP8
Quant 8 (Q8)
Quant 4 (Q4)
NF4
From my testing, I’ve noticed Quant 8 (Q8) offers slightly better quality than FP8, providing finer details while maintaining manageable resource requirements, but other works nice too. Pick the version that works best for your setup
Known Limitations
NSFW Capabilities: Still a weak area in this version. However, a minor fine-tune focusing specifically on NSFW content is already in the works.
Text Rendering: While text generation is better, occasional artifacts like odd symbols or incomplete words may still occur. But noticied usage of t5xxl fp16 instead of fp8 helps a lot with text
Tips for Optimal Results
Sampler: Use DPM++ 2M samplers for smooth and consistent outputs.
Steps: Aim for 30–50 steps to capture finer details without over-processing.
Scheduler: Beta Scheduler remains the best choice for this checkpoint.
Prompting TipsThe best prompting style involves complex prompts with clear, comma-separated phrases. While you can get creative with storytelling prompts, unnecessary descriptions like “this crap added more vintage to her style” won’t improve the results. Keep it concise and descriptive, focusing on essential visual details for the best output.
Future Plans
I’m committed to further developing this fine-tune. The next update will likely focus on:
Expanding NSFW capabilities
Enhancing edge cases like dynamic poses and lighting scenarios
Improving text rendering for sharper, more accurate results
P.S: If you still don't have realistic effect, then try add my ultrareal lora, usually helps me a lot
Ultra-Realistic Flux Fine-Tune v1
This is my first experiment in fine-tuning a checkpoint, built upon the foundations of my UltraReal LoRA and expanded with an extended dataset. The aim? To push realism to the next level, finding that sweet spot between amateur aesthetics and professional, high-quality visuals.
While this is only the first version and I see room for further refinement - the results are good, but not ideal (hands and feet can be broken sometimes, but situation is not critical, still better then defaul flux). This fine-tune isn’t just about amateur-quality outputs; it shines with professional-grade images, offering exceptional detail, lifelike shadows, and lighting. It’s a versatile model designed to unlock a wider range of realistic image generation possibilities.
This is very much a work in progress, and I’m sharing it to gather feedback and see how others use it creatively. If you test it, I’d love to hear your thoughts or see your results!
Also i uploaded both versions: fp16 (in ComfyUI it's better to use with e5m2) and fp8 and Q4_0
🌟 What’s New in This Fine-Tune?
Expanded Dataset: Nearly double the dataset size of the original LoRA, covering a diverse range of styles, lighting, and compositions.
Improved Realism: Sharper details, richer textures, and more natural lighting, bridging the gap between AI-generated and real-world imagery.
Versatility: From casual amateur-style snapshots to cinematic, professional-quality renders, this fine-tune adapts to a variety of creative needs.
Enhanced Anatomy: Better hands, limbs, and more natural poses compared to the base Flux model.
💡 Tips for Best Results
Use DPM++ 2M samplers for smooth and consistent outputs.
Aim for 30–50 steps for finer details without overdoing it.
Select the Beta Scheduler for optimal rendering performance.
⚡ Why Fine-Tune?
This fine-tune was crafted to overcome some of the limitations of the default Flux model. It enhances its ability to handle complex scenes while maintaining consistent quality across a range of prompts. The goal is simple: make ultra-realistic image generation accessible, reliable, and visually stunning, without requiring endless adjustments.
P.S: i plan to train this model more to make ultimate checkpoint with best anatomy and realism. This version is not very good with nsfw (this will be fixed in next version)
P.S.S: so far you can randomly get a low resolution image (dunno what exactly trigger this one, but will search for fixes). But seems like using high-resolution in prompt helps
Description
FAQ
Comments (127)
this is creating anime, plastic girls randomly ! never had it happen with v1
Dunno, I don't have any anime/3d/illustration images in dataset. Show me the prompt please, maybe I can help you
Thank you for your time and effort. I look forward to your next version even before I finish downloading this new one.
Thanks.
My results in Forge UI look nothing like this with the same promts. Any additional features you do in Comfy UI?
What about other settings? Like scheduler and sampler?
P.S: i'll install this Forge and check it by myself
okay, i installed, but can't understand what it doesn't want to load my checkpoint
@Danrisi You need to manually download and active VAE ae.safetensors clip_l.safetensors t5xxl_fp16.safetensors . I've got the same probleme minute ago lol
@lawiz2222149 yeah, I did it, but I don't have in top bar option to select vae / text encoder as it was shown on github
@Danrisi Please make sure to run update.bat, that at least made the vae / text encoder box appear for me. With Forge I'm getting at least reasonable results with DPM++ 2M / Beta.
@aplakka oh, thanx, need to try
Not sure if it's any help but I've found Euler simple sampler/scheduler at around 30 steps to be a bit more consistant with forge
@Danrisi I tried almost all sampler and scheduler combinations, mostly DPM++ 2M, Euler, or DPM2 with Simple/Normal/Beta. When I copy your node info from the results you post and paste it into Notepad++, it gives me like 10+ nodes you added in Comfy UI, so I assumed you did something additional, not simply a combination of your Fine Tune+Prompt+Guindance+Steps+Seed
Most striking difference is camera angle, especially in your "Elf woman with armor in a wood" prompt, dutch angle or similar camera shot never worked for me. But except angle, there are more differences, like skin color, posing. Like, obviously something if off
@yurigrigoryev309 i just solve a problem and was able to start Forge. I tried with fp8 and everything works good https://gyazo.com/e3a95b45ecaf3136ad2433fa0e2d6d1e
@Danrisi what about using the same seed and guidance? The exact resolution also matters
@yurigrigoryev309 as I remember same seed from comfy doesn't work properly with a1111 and vice versa
@Danrisi Makes sense, then. Can you please add some of your sofisticated works made with long prompts, like before, but made with Forge UI, so people without Comfy can test of your fine tune works properly?
@yurigrigoryev309 u mean just make the same, but in forge and upload in examples?
@yurigrigoryev309 made some examples in Forge, u can check images with ferrari and elf-lady
@Danrisi No, not necessarily the same prompts you did before. Just any complicated prompt making use of fine tuned style. I see examples, thanks
Could you do the full model fp16 instead of bf16 and the Quant 8 version please
Hi. But bf16 is almost the same as fp16. I trained in bf16 precision. Also what chatgpt says about the difference: "FP16 is a 16-bit floating point format typically used for GPUs.
BF16 is a newer 16-bit format often used in AI accelerators like TPUs, as it trades some precision in the mantissa for greater range."
Also i already have quant8 version, it's labeled pruned fp16 and u'll see gguf extension
@Danrisi ohh okay then thanks for responding, looking good so far , can't wait for you to add the nsfw recipe 😜
when I go on my workflow and supply change V1 into V2... its comes up with a bunch of errors
Hi. What kind of errors? Comfyui or something else?
@Danrisi I just understood the issue... I downloaded the wrong model lol
quick question If i downloaded on V1.0 the Fp16 22gb version...
Do I also download the V2.0 the Bf16 version?
is there a difference?
@Danrisi Okay! I fixed the issue... I downloaded the wrong model, thus why it was giving me a mismatch error.
thank you.
i didn't expect your NF4 version to work kinda good on Forge UI, mate you really outdone urself with this one.
again great work my dear friend, i appreciate your hard work!! 👌👏
Thanx for review 😀 Yeah, I too was surprised that the nf4 version turned out pretty good
@Danrisi @Danrisi yeah mate your NF4 is not bad for me, but after quite sometime of using it i noticed that it's not really that consistent with the results, like when i wanted to copy some the prompt of ur images you generated with forge ui (the human elf in armor one and the possessed Nun), it kinda gave me bad result (that default Ai-ish look and majority of the time it dosen't even follow the prompt), is that because of NF4 is not trained good unlike the other one's u did? 🤔
@MoudaH yeah, that's what happened with me in some scenes. sadly but there is no control when making nf4 version of other quants. i need to investigate this quaestion more, maybe will find some cure for this. thanx for testing and feedback
Great model!
I like it.
How about the results for Asian people?
Made for you some images with asian girls https://civitai.com/images/46205466
@Danrisi Thank you very much. It looks like it's working quite well.
I am getting this with every checkpoint loader I try:
ERROR: Could not detect model type of: D:\ComfyUI\models\checkpoints\ultrarealFineTune_v20.safetensors
Hi! You should place model into unet folder and use diffusion loader node instead of checkpoint loader node. cause this model doesn't have any baked clip or vae
do you know the best way to correct for this in SwarmUI?
@Fluffies yes. put in stable-diffusion loader. also u need clip-l and t5xxl in text_encoder and ae.safetensort in vae
no gguf ?
gguf yes. q8_0 and q4_K_M. look for pruned fp8 and pruned fp16
Hey Danrisi, about the v2, could u differentiate the name of the files, cause when we use wget, it seems to force only certain files to be downloaded, as they are all the same name, forexample i cant get the bf16 one via the wget function cause the name is similar to the other 2 safetensor files. appreciate the work btw!
Hey! Thanx for let me know, I'll fix it
[Suggestion]
Thank you for your work.
Can you please strengthen your support for Asians?
I'd like to see better representation of Koreans,
Japanese, and Chinese by age group.
how to download gguf version ?
Under "Details" field you will see "5 Files" click on it to expand. Then find pruned fp8 for quant 4_K and pruned fp16 for quant 8_0
Hello and thanks, I'm using your workflow, you suggest to add your Lora (already downloaded), where I have to insert in the workflow? Thank you.
i already have node LoraLoaderModelOnly (maybe under some image can be workflow without it) in workflow. u just need to enable it (or to add it. must be gguf loader or diffusion loader then our lora loader and then anythingeverywhere in front of this gguf loader). but i recommend using this lora with checkpoint only if u need really amateur quality, but it decrease quality
@Danrisi Hi, thanks for replying, I wish to use a Lora either for amateur quality or to use my own Lora for the face.
Another thing, I can't find any "CLIPTextEncodeFluxNUKE" node, I tried to search for it but nothing, it's missing even if doesn't stop the workflow. ...
@giusparsifal it's experimental node from this man https://www.reddit.com/r/StableDiffusion/comments/1ha76r3/new_text_encoder_clipsae_sparse_autoencoder/ ,but it's unnecessary in this workflow, i just tested something and didn't remove it
@Danrisi Oh ok thanks, I'll try to download it then, even if, as I said, the workflow seems to work fine even without it.
@giusparsifal you are welcome 😊
Which one is the Q8? (there are 2 x FP8)
Hi. Pruned fp16
what is the difference between flux 1.s and flux 1.d which models should I choose?
you mean when u select filters on civit? flux.s it's a light model (flux schnell). Flux.d - flux dev
@Danrisi thank you.
I guess, I should use schnell 'cause it's the free one and fit for personel usage.
Hello, im new to this and im using your workflow. I want to try out the pulid but im not sure how to connect it to the rest of the workflow
Hi. Pulid model connections must be between gguf model loader and anything anythinganywhere in front of gguf loader
Error: Failed to extract model file information from https://civitai.com/models/978314?modelVersionId=1164498 fix pls
Hi. Can you please clarify? It's comfyui's node? Cause I can't understand the source of problem and what I need to fix
@Danrisi yes and by comfyui
@2ef6t7j7nz Yeah, I understand you. I noticed that for some reason the full model doesn't share any metadata, which might be causing the issue, but the LoRA's metadata seems to be readable without problems. I trained both of them in the same place, so I'm not sure why this is happening
@Danrisi I use the models I download from the cloud on this site openart.ai the problem must be with the civitai site and I don't think it's with the model
some models from other creators cause this type of problem
the 6GB one... is that flux?
because i'm getting mismatch errors trying to load it.
If you're talking about the 6.24GB file, then yes, that's the Flux nf4 version, which requires a special loading node in ComfyUI. Personally, I don't like the quality of nf4. You might want to try the pruned FP8 (quantized q4_K_M) version, which weighs 6.46GB (if you are looking for something VRAM-friendly)
Windows fatal exception: access violation
need ur help,plz
Hi! Can you please clarify your problem? What app are you using? After what did this happen to you?
excelent model! any chance we can get better NSFW soon? :D
Thanx <3 Yeah, just right after The New Year will do the next training =)
Awesome model! What is your opinion on SD3.5? Thank you.
Thanks! 😊 It's hard to say much about SD3.5 since I haven’t used it extensively yet. However, I do love its aesthetics. That said, the anatomy still needs significant improvement—it’s been tricky for me to fix, even with LoRAs (I’ve trained some specifically for hands and feet, but they didn’t quite resolve the issues). Maybe I missed something during training, but for now, I’ve temporarily paused working with 3.5
hey , i keep getting blurry results even with recommended settings , can you please link the clip and ae and t5x files used alongside your flux
Hey. Blurry you mean low res and low quality or blurry you mean with blurry background. What about vae, and t5 i'm using default models. I just have custom clip-l from this man https://www.reddit.com/r/StableDiffusion/comments/1ha76r3/new_text_encoder_clipsae_sparse_autoencoder/
@Danrisi that fixed it , was getting blurry image , what do you recommend for cfg scale
@MobileOctopus744 2.5 - 3.0 guidance. also i get interesting result with cfg 0.8-0.9 instead of default 1.0
@Danrisi thank you love this model
@MobileOctopus744 u are welcome <3
after using this model for a couple of days i really love it best results at cfg 2 and 20 steps its almost as fast as sdxl for me , best results i got are with remastering old sdxl photos into flux to get the most detailed consistent results , i hope you figure out a way to make results more consistent sharp , out of a 12 batch 3-4 of them are good quality the others are either very low quality or blurry , and hope you figure nsfw part as well
@MobileOctopus744 Thank you for your detailed feedback =) I'm really glad to hear you're enjoying the model and finding good results with it at those settings.
I totally hear you about the consistency issues. Improving sharpness and overall quality across all outputs is one of my main priorities for the next update (but who knows if it will be fixed). Right now, it's a bit hit-or-miss in batches, but I'm experimenting with new methods to make the results more reliable.
As for NSFW content, I’m keeping it in mind as I refine the next version, so stay tuned! Thanks again for your patience and for sharing your experience 😊
I am getting this error pretty much as soon as I run a workflow
UNETLoader
Error(s) in loading state_dict for Flux: size mismatch for img_in.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([3072, 0]). size mismatch for time_in.in_layer.weight: copying a param with shape torch.Size([393216, 1]) from checkpoint, the shape in current model is torch.Size([3072, 256]). size mismatch for time_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072]). size mismatch for vector_in.in_layer.weight: copying a param with shape torch.Size([1179648, 1]) from checkpoint, the shape in current model is torch.Size([3072, 768]).
Comfy is up to date, I ran a git pull also to be sure. Not sure how to get around this one. My workflow works with other Flux models just not this new one.
This is quickly becoming my favorite flux fine tune. Its super easy to get realism out of it compared to other fine tunes on this website. It deserves more credit and it should be more popular. It works well with my loras and has better, clearer faces than other fine tunes i have been trying. Cant wait to see what you do in the next version. Awesome work!
Thank you so much for your kind words and support! 🙌 I'm really glad you're enjoying the fine-tune and finding it useful. The next version should be out soon (hopefully within a week or so, but I'm still experimenting).
I know some people have mentioned issues with blurred results or lower-quality images in certain cases. For now, a temporary fix I've found is combining it with my ultrareal LoRA at around 0.85 weight (though it might slightly affect anatomy). I'm working to address these issues in the new version, but as always, it's a bit of trial and error.
Thanks for your patience, and I'm excited to hear what you think of the update when it's ready! 😊
Getting the same error as many others (but can't see any mention of a solution)
UNETLoader
Error(s) in loading state_dict for Flux: size mismatch for img_in.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([3072, 0])....
Hi. I just see one man asked me yesterday about this. Yeah, and the temporary solution was to use gguf loader with gguf model instead of using fp16 or fp8 version in diffusion loader. Honestly dunno what's wrong, cause everything is working good for me (i'm using comfyui updated a few weeks ago), and i think problem is with new updates
@Danrisi OK thanks for the input - it's really weird, I don't have issues with other models, but will certainly try the GGUF version
@Riotsurfer btw. i downloaded again my bf16 version of checkpoint and used diffusion loader with fp8_e4m3fn and everything works good. So the problem in comfyUI updates.
P.S: I have my suspicions as to why other people's models work fine and mine doesn't. I think it's because i trained in bf16 precision, not fp16
love the examples people are making with this, but I'm only used to gguf files, am I supposed to put this into checkpoint folder?
Hey. If you are using Comfy then you need to put in Unet folder, if something else then put in folder stable-diffusion
@Danrisi ohh it works now. can't wait to play with it on the weekend. thank you so much for your work dan!
I download these checkpoints but can't run because of this error:
"AssertionError: You do not have CLIP state dict!"
Hello! This is what users wrote about this problem on reddit:
"This means you need to load additional modules not included with the checkpoint (they’re not baked in). Forge has an annoying bug where, if this happens, you need to successfully load any other model before trying to load it again. I submitted a PR 2 months ago that resolves this but Illyasviel needs to review it and they are likely too focused on some other new groundbreaking project" (c)
@Danrisi This is what you do if you're running from github diffuser library as I have done. I have a script in my project to do this now that I know what Civitai d/l really means.
from diffusers import FluxTransformer2DModel
from optimum.quanto import freeze, qint8, quantize, quantization_map
from pathlib import Path
import json
################## FLUX QUANTIZE TRANSFORMER FROM SINGLE FILE #########################
base_model = 'C:/Users/xxxxx/.cache/huggingface/hub/Civitai Models/ultrarealFineTune_v20.safetensors'
dtype = torch.bfloat16
transformer = FluxTransformer2DModel.from_single_file(base_model, subfolder= "transformer", torch_dtype=dtype)
quantize(transformer, weights=qint8)
freeze(transformer)
save_directory = "./flux-dev/ultrarealFineTune/fluxtransformer2dmodel_qint8"
transformer.save_pretrained(save_directory)
qmap_name = Path(save_directory, "quanto_qmap.json" )
qmap = quantization_map(transformer)
with open (qmap_name, "w" , encoding= "utf8" ) as f:
json.dump(qmap, f, indent= 4 )
print('Transformer done')
return
Incredible Model. Fascinating stuff. Can't wait for v3
Thanks! 😊 I just need a bit more time to test my new datasets on LoRAs before I can start training the V3 checkpoint. Plans for V3 include better NSFW support (improved details for female anatomy like genitalia and breasts), enhanced lower body anatomy (legs, feet, toes), improved text generation, and overall better textures
@Danrisi I can't wait. That's perfect.
I think for those downloading files from this site that it is made more clear that you're downloading a Transformer and not the full model itself. Big difference.
Having said that, I'm giving it a go right now, so far, images looking very good. Very impressed. Using a QINT8 version of this model quantized from the BF16 d/l. FWIW I am using prompts from Sana:
Thank you so much for your feedback and for taking the time to test the model =)
I wanted to clarify something regarding your comment about "Transformers vs. Full Model." Could you explain a bit more about what you mean by this distinction?
Thanks again for your detailed comment and for sharing the prompts you're using. I truly appreciate your support ❤️
@Danrisi My comment is directed at Civitai and not you. Your model seems to be very good from what I've seen. As for the comment on Civitai, the d/l is basically the Transformer folder of a "Full Model" and as such when I run it I can't use the normal model.from.single.file(blah blah) that you normally do when running diffusers. Once I figured that out I did my normal thing to get it running with no problems. BTW, keep up the good work. I know the time, effort and expense of doing this is consuming, but advances what is otherwise a pretty good base model.
"From my testing, I’ve noticed Quant 8 (Q8) offers slightly better quality than FP8, providing finer details while maintaining manageable resource requirements, but other works nice too. Pick the version that works best for your setup"
I don't have the VRAM to run a full model in BF16 mode even though I have a RTX4090. The QINT8 version seems to be very good. I don't bother with FP8/FP16.
And the images I posted here were run w/o using any Loras. I'm using a home grown scheduler for Flow Match since the diffuser github people are slow in getting these schedulers / samplers out.
@condzero1950 btw, i'm using q8 with my 3090 too
@Danrisi So you're FortranUA on Reddit, eh? I'm NoSuggestion6629.
@condzero1950 yeah =) oh, i remember you
@Danrisi I can be grating at times. :-)
This model is very nice, but can you make it produce good images with a small number of sampling steps? Maybe 20 or less than that
Not sure if this possible to do. Maybe just to overtrain a model
@Danrisi i'm requesting the same but with only 3~4 steps ...
@amazingbeauty OP can't magically do this. The number of steps is limited by the design of the Flux model this based on.
If you want fewer steps, try SDXL lighting, which is amazing with only a few steps, and not much worse that Flux.
Again, this isn't really something the OP would want to do. The number of steps is related to the design of the underlying model (Flux).
Check out SDXL Lighting. It's almost as good as Flux in most cases and there's tons of cool Loras already.
@DetroitArtDude SDXL Lighting exists - me using SDXL non-Lighting with 60 steps for more quality 😏
I think it's just a question of quality or rapidness, someone wants fasterr gen for 10 sec, someone can wait 10mins for 1 image and okay with it
You can try this Lora https://huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha
At weight 1.0 with Euler Simple or Beta and get nice generations with 4~8 steps
@DetroitArtDude WTF you are talking about, FLUX DEV require 20 - 30 steps (not 50) to get excellent results.
@zerocool22 for illustrations maybe yes. but realistic images better to generate with more steps and not with euler
Hello,
I am getting this error
Model in folder 'vae' with filename 'FLUX1\ae.safetensors' not found.
where do I download the approapriate VAE file?
Hello there. Sorry for not responding on reddit. You need to download vae from huggingface https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/main/ae.safetensors and place it to vae folder
This is a basic diffusion setup issue. It's not the OP's problem that you're getting this error. This space is for constructive feedback / commentary on the model. There's dozens and dozens of good Stable Diffusion / Flux, etc tutorials out there. You are better off reading those than posting on here. Best of luck with your generations.
@DetroitArtDude The developer answered so nicely above, and you come here to complain and hate.
Who are you, anyway? Did you contribute to anything?
LOVE, my friend! Spread it, and you will receive it back. Hate works the same way.
@Danrisi Thank you for your attention and caring. It solved my problem.
Hi, I don't know why but it works only on Forge. In ComfyUI I always got error.
Hi. Tell me more about the problem. It's with nf4/gguf/fp8 version?
@Danrisi Hi, the .gguf work, the .safetensors doesn't.
Another question, finally the NUKE node works but, in your workflow I don't understand how to use it. Maybe I have to replace CLIP Text Encode (Prompt)?
Thanks anyway!
@giusparsifal what exactly wrong with .safetensors? cause for me it's work in comfy via diffusionloader node. what about nuke node, you don't need it (just remove it or disable). it was just for experiments, but if you want too, then you can read an article and guide from the author https://www.reddit.com/r/StableDiffusion/comments/1ha76r3/new_text_encoder_clipsae_sparse_autoencoder/
@Danrisi A short version, it'very long, of hte error I got:
UNETLoader
Error(s) in loading state_dict for Flux: size mismatch for img_in.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([3072, 0]). size mismatch for time_in.in_layer.weight: copying a param with shape torch.Size([393216, 1]) from checkpoint, the shape in current model is torch.Size([3072, 256]). size mismatch for time_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072])
Details
Files
ultrarealFineTune_v20.safetensors
Mirrors
ultrarealFineTune_v20.gguf
Mirrors
Available On (4 platforms)
Same model published on other platforms. May have additional downloads or version variants.