@Danrisi Alright thanks alot

@Danrisi Thanks imma use this right now i already used the v2 version a few times its really cool

@Danrisi Any chance of an fp8 safetensors?

@Danrisi epic!

@Danrisi Would it be possible to see the pruned FP16 as exl2 in the future? My GPU can fit the FP16 versions fully but GGUF is a little slower since it's more optimized for CPU :)

My Forge always crashes when using this model with loras. Don't know why.

@Danrisi It's weird. Always connection timeout and ui freeze after 94% of generation. Tried with diff Samplers+Schedulers. I think my system is running out of memory, when using your model (Full Model fp16)+ a lora. Still wondering why, I have a good PC here, 64GB DDR5, GeForce 4080. Yes, try it yourself ;) You may find a reason. Thank you for your work and effort!

Edit: everything is working fine with the fp8 model. Guess forge + your fp16 model is cooking my PC. Cheers!

couldn't get good results with both fp8 and the gguf , both look much worse for me than v2 , tried 50 steps and cfg 3 ,v4 lost alot of detail and realism

@Danrisi here is a comparison , same prompt and loras , v2 at 20 steps , v4 at 30 , if i set it to lowers steps i get worse results https://i.ibb.co/27WwYqHY/00100-3366942652.jpg

using cfg 3.5 gives better results , but your lora UltraRealPhoto needs an update it destroys the quality of v4 at anything above 0.4 compared to v2 where i use it at 1.5 and get great results

@Danrisi understandable , the problem is the loss of detail with this v4 , that lora added alot to the skin detail and overall authenticity , v4 with amplifier lora looks soft and has that ai fake feel , hope you can improve it

@Danrisi please consider adding the 6 gb nf4

@Danrisi thanks, glad for the fp8 option!

the gguf also goes into the unet folder.

@kakkkarot Thank you!

@Danrisi thank you, only been at this a few days so i appreciate people explaining to me the nooby stuff

I've heard that SUPIR is da proverbial bomb, but it's not as plug and play as ultimateSD—it requires comfy.

https://github.com/Fanghua-Yu/SUPIR ← OG Github

https://civitai.com/articles/6180/easy-guide-how-to-improve-your-images-with-ai-tools ← Comparison (beware of dead links)

Thanks @Lathos I tried some upscalers with this model but they mostly introduce that flux background/plastic skin back in. Trying now different MP settings. @Danrisi if you don't mind, what resolution/aspect ratio were the most pictures used for training? I got the 2000s look with 2:3 1MP, a more 2010s look with 5:7 1MP and a combination of both which makes the background look more real and well rationed with 16:9 1MP. Increasing the MP in general makes the person look imposed on top of the background. The issue with the aspect ratio is that if the prompt is too long or contains uncommon details flux will hallucinate and try to fit in too much in background.

@MonarchTrayner Yeah, that incurs on what @Danrisi said about it losing the aesthetic quality upon upscaling. As far as I know, SUPIR uses an SDXL model in its upscaling process, so if you're willing to give it another chance, you could perhaps find an SDXL that somewhat matches the aesthetics of ultrareal. UltimateSD Up has the advantage of allowing for FLUX models, I think.

https://www.youtube.com/watch?v=EMAz8KktB5U&t

@Danrisi To be honest, I feel that low res compliments the amateurish style... This definitely isn't me coping about my troglodyte gpu.

But yeah, SUPIR is basically a modified controlnet model, it will always struggle with finer aesthetic details like this because it's using a whole other model to regenerate the image and interpret details. Maybe one day they'll release a flux version of it and it will be possible to use UltraRealFT.

I merged this model with Lumiere Alpha from Aixonlab

https://civitai.com/models/966796?modelVersionId=1082472

did a 50/50 merge. Results have been pretty good with more detail. Better than what I think you'd get by adding just a lora.

@condzero1950 Are u considering to publish or share ur weights?.

@AndSudo02 I have a 23 GB FP16 safetensor file that I can upload to anyone who wants it.

@condzero1950 It would be great if u upload it to HF, there not a lot cloud services that give 25Gb of storage for sending it privately ):

@AndSudo02 I have uploaded this file to hugging face. You can find it here: https://huggingface.co/condzero/flux1d-merge-model/tree/main

download the safetensor file and use as you would normally use any civitai model file.

Let me know what you think.

@Danrisi I have uploaded this file to hugging face. You can find it here: https://huggingface.co/condzero/flux1d-merge-model/tree/main

download the safetensor file and use as you would normally use any civitai model file.

Let me know what you think.

@condzero1950 I aprecciate it man, im about to download the weights, i´ll report back with the results .

What I'm finding is that if you mix "Good" Loras with certain models, it can produce better results (of course this is subjective in nature). I have 3 models setup; (2) base Flux with "Realistic" Loras and "Artistic" Loras and 1 Ultra-Real fine tune with "Realistic" Loras.

Base Flux 1.D Realistic Loras:

<lora:aidmaFLUXPro1.1-FLUX-v0.3:0.7>aidmafluxpro1.1

<lora:aidmaHyperrealism-FLUX-v0.3:0.7>aidmaHyperrealism

<lora:FLUX-dev-lora-add_details:0.7>skakker labs

Base Flux 1.D Artistic Loras:

<lora:aidmaMJ6.1-FLUX-v0.5:0.7>aidmamj6.1

<lora:flux_vividizer:0.7>

<lora:MJ6.1_Flux-PRO:0.7>

You may want to try this and see what you think. Strengths are all 0.7 which seemed like a good compromise.

@condzero1950 I was comming to say something similar, even with the regular Euler/Beta/25ish steps, it gets LoRAs rlly well, i tried with some i had with bad training or just were trained at low resolutions, manages to fix artifacts. Overall, has a nice balance between prompt adherence and creative results.
My best results are with IPNDM sampler, beta, 25 steps, maybe adding a lora that ads more details to the face would put this on the top finetunnings, thanks for sharing, it will be my defalut model for now (:

@Danrisi excellent short answer , step has same speed all the time with same model (fkux or etc) and same H.W.

Should have said - I'm on ForgeUI.

@redlittlerabbit 14 days later... I only saw this now. I had the same problem when I switched to forge, turns out it was the T5XXL model I was using, so play around with that. I'd point you to the t5 model I'm using, but I don't recall where I got it. Make sure to also put the CFG scale to 1 and only play with the Distilled CFG scale slider.

upd i found the lora!

@Danrisi Yes! It worked, thank you so much!

Hi, were can I found this workflow?

good luck bro, I been working on ts for probably 7 days and today I made it work here, I'm a noob to ts but Chatgpt helped me, try working with chatgpt to make it work

@Danrisi its totally okay bro this isnt your job, you are helping people already! And it was good for me to find my way I learned a lot in the process. But since you answered, do you know @catsoupai from instagram? any idea how he create those images? if you dont know, you should check it out

i know how to use it. the problem is that i can use any model except this one

Danrisi's workflow is simple and solid; you can get it by clicking on one of the showcase images, clicking on workflow:nodes and ctrl+V-ing it on your comfyui canvas. You'll need a few custom node packs, but those are trivial to install using ComfyUI manager. Alternatively, the workflow can be used with no custom nodes by just doing a few tweaks.

ty g

@Danrisi Thanks, i will try to use ComfyUi

@Danrisi both, basically just seeing what it likes best. Because I've been using it at just 1024x768 and it seems quite well, but just curious if it's got some more spice left in it for higher resolutions lol.

@Danrisi Didn't really understood, Where do I find the Pruned FP16 model for the gguf? Thanks!

@endlessblink I just discovered that under the "Details" tab there is a "files" tab.

@mazurenckoslavik377 Thank you!

Hi !

Thanks for your reply, i tried to upgrade in the last version 0.3.26 and it doesnt work. I have a 3060 Ti 8GB, i downloaded the Full Model fp8 (and other). The ultrarealFineTune_v4.safetensors is in models/unet. And when i select to drag drog the node it doesnt do anything. I dont know why, i must search, perhaps i must install some custom node from custom manager at this point to execute it.

@Danrisi i dont need shcnell i think i really need nvidia 4090 GPU (Y) . not to bother about running flux on my old cpu again that painfully lost of time . keep the good work.

Wow! I had no idea I could actually do that!

thanks

@Danrisi thanks will try that

@Danrisi doesnt seem to recognise the file as being there

@Danrisi nevermind I fixed it

@Danrisi ComfyUI. Two Load Checkpoint nodes into a ModelMergeSimple node and then simply ModelSave node

I have the same error. Were you able to fix it?

@1de666vil3458 no, still having that error..old my old download models work..this is problem wd new ones

@Danrisi Thanks for the super quick reply! I had been trying this, but with CLIP loader, not dual. It works now, but with the CLIP I found online it seems to generate very dark images, sometimes just black.

Which CLIP file should I be using?

@Danrisi thanks for the answer, maybe nf4 won't help me, the thing is that I want to use flux model in krita ai diffusion, but for some reason flux models don't work there, considering that I installed text-encoder and vae in all the corresponding folders, if you are knowing krita ai diffusion and can you help me with the correct installation?

Not all heroes wear capes.

you just solved my problem with a snap of a finger (before that I worked on it for half a day)

No. Flux doesn't work on foocus

Which version should we get? Does this include Forge?

I've gotten good results with 20 steps for reasonable-ish time on my paleolithic 2080Ti, the 50 that @Danrisi sugges is, well, a suggestion.

@Danrisi nono, just a Lora for it lol

@Danrisi ohh okay, nevermind thanks , cant wait for v5 tho🙌🏻

@Danrisi Is it good for fine-tuning a character? Will that affect the realism?

@Danrisi i was just wondering which model you generated the showcase images with

If you haven't fixed it yet, I have been using UltralyticsDetectorProvider instead of MediaPipe node as a workaround.

Use SwarmUI instead of Comfy

@Skandinaavlane  Thanks, you know I had the same error and it was driving me crazy. In the end, I did what you said — I replaced it with that one and now it works 👌

Open any image made by the model author on Civitai and copy as usual. Then just Ctrl+V

same

Those samples have been created using the "Realistic Amplifier for UltraReal Fine-Tune" LoRa. I copied the seed and prompt from the image of the lady holding the beer bottle, used the LoRa at strength 0.7 and recreated a similar good looking image in just 28 steps.

But sure, it might be better providing those samples without any LoRa.

Danrisi Question: How does it fare with character loras? Currently the Q8 version with just my character lora is creating undesireable results, the character is inconsistent, face looking off. If I add in your realistic amplifier lora will that fix it? Do I need to retrain my character LoRa on your checkpoint?

Full Model fp8 (11.08 GB) SafeTensor = fp8

Pruned Model fp8 (6.46 GB) GGUF = Q4_K_M

Pruned Model fp16 (11.85 GB) GGUF = Q8_0

Full Model fp16 (22.17 GB) SafeTensor = bf16

@Il_ya thank you!

@Danrisi too real looking, they were suggestive in content and they some were flagged as being unable to determine if they were AI generated. I can't see from the blocked labels why they were blocked.

i have 8gb vram, 2-3 min

20-60s (Depends on settings & LoRAs)

You're obviously doing something wrong that's pushing you way over the VRAM limit. Scale the resolution & step count back until you're under 1min generation time. Wouldn't hurt if you have at least 32GB of RAM as well, but it's mostly really just VRAM (& Tensors) that's important with image generation.

Did you download the Full Model fp16 (22.17 GB)? Because that's suited more for a 24GB VRAM GPU like a 4090 btw. If you can't fit the entire model into VRAM, then it will really cripple your generation speeds. (Also, the size of the checkpoint/model, doesn't always mean you need that much VRAM necessarily either, but it's a decent rule of thumb if you're starting out).

(Also, I'm going to assume you're new, but you must set CFG Scale to 1 when using FLUX). Try testing a resolution of 1024x768, 20 steps. That should be under 30s. If it isn't, then something else isn't configured right.

Upload one of these "5-7 minutes" image generations in your profile gallery as a .PNG file if you want. I (or someone else) can easily take a look at the metadata to see where it's going wrong (if it's a bad setting).

@Danrisi yes, i usually use fp8 (11gb) and works fine, but now im trying gguf (for better quality)

(Just for anyone else reading & wondering) GGUF models are generally less accurate & inferior to full FP8 btw (most are quantized for lower end setups), & contrary to somewhat popular belief, are actually typically intrinsically slower with the added overhead, too. They will only be faster on setups with insufficient VRAM to use something else - which will cripple the generation speed as it offloads to RAM instead if there's not enough VRAM to use.

https://civitai.com/models/691446/bytedance-hyper-flux-acceleration-lora (This may be useful if you're on more modest hardware).

@ShadowCell Hi, I have a question: Is there much of a quality difference between FP8 and basic F16? I use F16 FluxD with the T5xxlfp16 in Forge... Obviously, I've tried this checkpoint and I like it, but I notice it's at its limit and takes a long time when I use hiresfix or generate directly with 1920x1080. / I have a 3090 and 32 GB of RAM.

@ShadowCell this is not fully correct Q8 is slightly better than fp8 that's the result of my tests. As for speed I recommend using the multigpu gguf node. Generation speed on a 4070 12 GB and 64GB Ram is fine even when using 18GB Q8 checkpoints.

Well @djramonx811, there is not a huge visual quality difference between FP8 and FP16 for most AI image generation tasks, especially with models like Flux.1D or T5XXL in Forge. FP16 is the “safe default” for stability and compatibility, but FP8 is much faster if your hardware and the model support it (which in your case, does not). The RTX 3090 does not have native FP8 support. It is only available on the newer Ada/Hopper cards (RTX 5000/6000 Ada or H100, for example). So even if Forge offers FP8 as an option, your 3090 is still running everything in FP16 or FP32 under the hood, and that means you will not see the performance boost from FP8 on your current card.

Now, some people asked if there is an issue when running FP8 models on an RTX 3090, and the answer is yes. PyTorch and CUDA will “emulate” FP8 using FP16 or FP32 ops, which can be slower than a true FP16 model and is definitely slower than running FP8 on a 4090/5000/6000 or Hopper. If you see “FP8” in a model’s logs on a 3090, just know it is a compatibility thing. No speedup, no VRAM drop, nothing bad happens -just do not expect magic.

Also, The slowness at high resolutions is just due to VRAM limits and the heavy compute load at 1920x1080 with Hires Fix (which I do not recommend using with FLUX due to those common vertical lines).

If you want the fastest, most stable Forge/FLUX.1D setup on RTX 3090, just use:
--xformers --cuda-malloc --opt-sdp-attention --opt-channelslast
All other flags are optional and do not improve speed or stability. (Unless you enable Triton) That gives me 3.15–3.26s per-steps time, or 1:02–1:03 for 20 steps.

But wait, for even more speed. I changed all of these settings to get 2.81–2.88s per-step time, or 0:56–0:57 for 20 steps;

1. Always Discard Next-to-Last Sigma

Default: Unchecked

Current Setting: Checked

Function: When enabled, this setting helps to smooth out the final image by discarding the next-to-last sigma value during the sampling process. This can lead to a more refined output, especially in the final steps of generation.

2. Negative Guidance Minimum Sigma

Default: 0

Current Setting: 3.0

Function: This parameter allows you to skip applying negative prompts for certain steps when the image is nearly complete. A higher value can speed up the generation process by reducing the influence of negative prompts, which can be beneficial for achieving faster results.

3. Token Merging Ratio

Default: 0

Current Setting: 0.1

Function: This setting merges redundant tokens during the forward pass, which can enhance processing speed. A higher ratio can lead to faster generation times but may also affect the quality of the output.

4. Token Merging Ratio for img2img

Default: 0

Current Setting: 0.1

Function: Similar to the general token merging ratio, this applies specifically to image-to-image generation. It overrides the general setting if non-zero, allowing for optimized processing in img2img tasks.

5. Token Merging Ratio for High-Res Pass

Default: 0

Current Setting: 0.1

Function: This setting is specifically for high-resolution passes and works similarly to the other merging ratios, enhancing speed while generating high-res images.

6. SGM Noise Multiplier

Default: Checked

Current Setting: Unchecked

Function: This parameter matches the initial noise to the official SDXL implementation. It’s primarily useful for reproducing images consistently. Disabling it may lead to variations in output.

7. Negative Guidance Minimum Sigma All Steps

Default: Unchecked

Current Setting: Checked

Function: When enabled, this setting allows the negative guidance minimum sigma to skip all steps instead of just every other one. This can significantly speed up the generation process but may impact the final image quality.

8. Eta Noise Seed Delta (ENSD)

Default: 0

Current Setting: 31337

Function: This parameter adds a specific number to your seed during image generation, which can help replicate results from other models or implementations. It’s particularly useful for achieving consistent outputs.

9. Prompt Word Wrap Length Limit

Default: 20

Current Setting: 74

Function: This setting controls how prompts are wrapped in tokens. If a prompt is shorter than the specified limit, it will be moved to the next chunk if it exceeds the 75-token limit. This can help manage longer prompts more effectively.

I hope this helps! ♥

@Hyokkuda 
I have an RTX 4090, does that mean I should download the FP8 not the F16 ?

Sorry for the wall of text, I am going go try and make things as easy to understand as possible and as simple as possible. ♥

@amrpp79697, the RTX 4090 supports FP16, BF16, and FP8 (hardware-wise), but FP8 is still pretty new and not widely supported across all models and frameworks yet (or so I read). If a model is labeled as “FP8”, that means the model was saved and intended to be run using FP8 precision. The model expects the underlying software and hardware to support FP8 operations. It may not run correctly (or at all, usually more affected on AI video rather than images) if you load it in a pipeline that only supports FP16 or FP32.

FP16 = works everywhere, best choice for now.

FP8 = only use if the model specifically says it supports it, and you know what you are doing.

For image generation, you generally do not need to worry much about FP8 vs FP16 unless your graphics card is underpowered, you are planning on generating a ton of images in a batch, or you are extremely impatient about waiting a few extra seconds per image. For most people with an RTX 4090, FP16 is more than fast enough and is way more compatible right now. So try not to think about it too much.

⚠️ FP8, FP16, BF16, FP32, etc... only matters for AI video generation. No exception!
If you run the wrong "precision" model for your GPU, your AI video generation will fail or crash!
If this is too much for you, just get .GUFF models for videos, because those are determined automatically by the software (based on what your GPU supports), or sometimes fixed by the author in the code which you can edit. Edit: RTX 3000 series (like the 3090) can run some FP8 models if they are in the "E4M3FN" format. In these cases, the model is quantized, and the inference libraries automatically upcast FP8 weights to FP16 or FP32 behind the scenes. This avoids crashes or errors mid-generation (which is a good thing!), but does not give any speed or VRAM advantage on those GPUs. However, sometimes it is actually better to use these FP8 "E4M3FN" models rather than good old FP16, since FP16 can freeze or hang if your VRAM is completely maxed out during the process. The FP8 model (via upcasting) can let the generation finish, where FP16 might just stall forever.

On an RTX 4070 12GB like @nimchik1439, this generation time is to be expected for FLUX, unfortunately. It is just a really heavy model. There are not any truly lightweight versions of FLUX available. FLUX models are designed to push the limits on image quality and flexibility, but that comes at the cost of much higher VRAM and processing requirements. There is not a “Tiny FLUX” or “FLUX Lite” variant.

The recommended GPUs for FLUX should be the following; RTX 4090, 4080, 4070 Ti, 3090, 3090 Ti, 4080, 6000 Ada. You will get reasonable generation times (20-60 seconds per image or so, at 1024x1024, FP16).

After that, the RTX 4070 (12GB), RTX 3080 12GB, RTX 3080 Ti, etc... these cards can run FLUX models, but you will often be right at the VRAM limit for high resolutions, and generation times can be 1–5 minutes per image, depending on settings. :S

And I really advice against using the RTX 2000 series or the RTX 3050(?), do I have it right? Stupid GPU naming... lol, Those cards will never cut it.

Rule of thumb (or whatever):
- 16GB+ VRAM is “ideal” for FLUX.

- 12GB VRAM can work if you are patient and careful with settings.

- Below 12GB only small images, with patience, and be ready for errors/crashes or artifacts once the VRAM kick the bucket.

I hope this was helpful to you and everyone else here. ♥

@Hyokkuda 
thanks a lot for your detailed reply, much appreciated

@amrpp79697 No problem! :3 I am always happy to help whenever I can!

@Danrisi I ran Flux GGUF models with 6 GB before my PC upgrade. It was slow but it worked with ComfyUI.

@Hyokkuda Thanks for the info and taking your time.👌

Comfy ui

@Danrisi I think the I found the issue, in the workflow, the advance k sample node is bogging down the generation. I have not found a solution, yet.

@Noob_ee yes exactly, same issue here

@Danrisi I do... with my 3090 and I liked the results... but it takes a while, especialy generating directly at 1920x1080. Checking the Stability Matrix console... I see a -634mb, that shouldn't be good?.

ALL Loaded to GPU.

Moving model(s) has taken 21.58 seconds

260 [Unload] Trying to free 1024.66 MB for cuda:0 with @ models keep loaded... Current free memory is 22984.49 MB Jol Cleanup minimal inference nemory. Done.

702 tiled upscale: 160% 24/24 [90:64409:00, 5.86it/s)

43 [Unload] Trying to free 1824.60 MB for cuda:0 with keep Loaded ... Current free memory is 22982.49 MB The Cleanup minimal inference memory. models Done.

765 tiled upscale: 100% | 24/24 [50:02408:00, 11.80it/s] 706 [Unload] Trying to free 7171.39 MB for cuda:6 with 1 models

keep Loaded ... Current free monory is 22897.26 MB ... Done. 260 [Unload] Trying to free 7171.39 MB for cuda:0 with 1 models keep Loaded... Current free memory is 22895.64 MB... Done.

768 [Unload] Trying to free 34797.86 MB for cuda:0 with 6 models keep loaded... Current free memory is 22927.13 MB nodel IntegratedAutoencoderKL Done. Unload

765 [Memory Management] Target: Kłodel, Free GPU: 23089.23 MB, Model Require: 22768.13 MB, Previously Loaded: 0.06 MB, Inference Require: 1024.00 MB, Remaining: -634.90 MB, CPU Swap Loaded (blocked method): 2616.00 MB, GPU Loaded: 26684.13 MB

Moving model(s) has taken 14.64 seconds

85s/it] 12/30 [83:19 46:40, 100.02s/it]5<2:32:18, 62.

@Danrisi I can't provide a screenshot via this comment (using Forge). Can I send a screenshot somewhere else?

@Danrisi v4 fp8 not 11gb it download 22gb fp16 , fix and upload the fp8 link pls

@Danrisi Yeah. I just wanted to make a "Back to the Future" joke...

@Danrisi what your opinion , it will lean to nudity or it's fine model better than others for realistic images ? ' i tested few dev fp8 realistic models and it take prompts that i have tested before with shuttle , take it to lean to more nudity'

I cant figure this out as well. I keep seeing this is not a base checkpoint and all the workflows use some gguf version. The creator is missing allot of key details on this.

Danrisi can i install any of hose clips you linked?

did you find it?

did you find it?

I am using OstrisAI. Can you tell me how to upload and train over your checkpoint on it? it doesn't show any option there.

@Danrisi Which platform you recommend to train Lora on?

@Danrisi thanks, ended up training my own loras, found out controlnet is completely uselless and doesn´t work in forge lmao.

You made absolute masterpiece tho good job dude fr, thank you!

Btw quick question how should I set cfg and cfg distilled I know you mentioned cfg should be 3 but what about sfg distilled?

Also one more question, for some reason I am unable to generate full body shots, could you please recommend the cfg + prompting

with the full 22 gb model?

With how many steps and what resolution? You probably should use the Q4 quantized version and then combine it with the Hyper8 LORA @0.125 strength. Works nicely.

@Danrisi Maybe Z-Image is also a good option

hey danrisi sorry if this is a stupid question but would you ever release a lora like this for SDXL based models?

@Danrisi ahh what a shame, well i will be looking forward to it if you ever do. your models and loras are truly top notch man, keep up the great work

@Danrisi Awaiting your ZIT checkpoint!

you might be running the wrong encoders (vae, clip_l and t5xxl)

download the official flux ones from blackforest huggingface

make sure you follow the prescribed settings (dpmpp2m + beta, 30-50 steps, flux guidance at 3.00)

should be one of these issues

@10sor I realize that there are a lot of people who ask questions before even reading the instructions but yeah I definitely made sure I ran all the recommended settings. But alas I can't get it to look like the example images or very realistic compared to the other models I use. The Digicam lora is fantastic with my other models though.

@aedrenavendale291 we are all people who sometimes ask questions before we read , its no big deal, what i would try if i were you is try some other models, browse to the community artwork tab, find a post where the prompt and settings are shared, throw these into your own workflow and try to get a 1:1 result of the original post. doesnt have to be with girls or anything, this really helped me to get familiar with copying results (not for the sake of copying ofc) but for reproducing or rather calibrating the workflow

You cant. I recommend comfyui for beginners.