V4.0-Full Version:
V4.0 模型(纯净基础模型)，是在 V3.0 的基础上，融合了 SRPO 模型的写实与细腻，Krea 模型的艺术性与多风格，以及良好的质感与LoRA兼容性，模型的综合能力有较大提升。
非常的真实与细腻，即使 TTP 直接放大至 8M 像素，依然能保持极好的细节。
提示词还原能力较大提升，推荐使用 LLM 强化和详细描述的结构化提示词。
良好的艺术表现力和风格多样性，良好的 LoRA 兼容性，包括 NSFW / SFW 各种 LoRAs。
The V4.0 model(Pure base model), built on the foundation of V3.0, integrates the realism and delicacy of the SRPO model, the artistry and multi-style capabilities of the Krea model, as well as good texture and LoRA compatibility, resulting in a significant improvement in the model's overall capabilities.
Extremely realistic and delicate, maintaining excellent details even when TTP is directly enlarged to 8M pixels.
Greatly improved prompt restoration capability, it is recommended to use LLM-enhanced or structured prompts.
Good artistic expression and style diversity, good LoRA compatibility, include all kind NSFW / SFW LoRAs.
Also on: Huggingface.co, Modelscope
===========================================================================
V3.0-Krea Version:
Flux.1-Dev-Krea 模型改善了 Dev 模型的艺术风格与写实摄影的能力，但人像的清晰度和美学方面有所减弱，特别是与原 Dev 模型训练的 Lora 兼容性很差。本 V3.0-Krea 保留了 Krea 模型的主要长处，改善了图像清晰度及与原 Dev 模型 Lora 的兼容性，但 Lora 兼容性方面改善不多(主要指人像和风格LoRA，其它LoRA及CN还行)，不太理想，这是该版本比较遗憾的地方，请大家慎重下载。
Flux.1-Dev-Krea has improved the artistic style and realistic photography ability of the Dev version model, but the clarity and aesthetics of portraits have weakened, especially with poor compatibility with Lora trained on the original Dev model. The V3.0-Krea retains the main features of the Krea model, improves image clarity, and enhances compatibility with the Lora, but the Lora compatibility is improved minimal and not ideal(Mainly refers to portrait and style LoRA, other LoRA and CN are acceptable), which is a disappointing aspect of this version. Please download cautiously.
推荐使用 GNER-T5-XXL 代替 T5-XXL，获得更好的提示词理解能力，你可以从 https://civarchive.com/models/1888454 下载，或我的 HF Repo 下载。
Recommended to use GNER-T5-XXL instead of T5-XXL for better prompt understanding capabilities, you can download it from https://civarchive.com/models/1888454 or my HF Repo.
一些示例图片 (Some example image) :
模型使用：
基本组合：deis+simple / euler+beta，您可以尝试不同的组合。
Basic: deis+simple / euler+beta, You can try more different combinations.
Also on Huggingface.co
===========================================================================
V3.0-PAP Version:
v3.0-PAP：人像与艺术摄影出图优化基础模型。模型在构图、光影以及东方人脸型等方面进行了专门优化，进一步加强脸模的敏感性、适配性。
与 Dev 原版模型相比，该版本在人种和脸型方面更真实、更丰富，脸模的例图来自以下作者的 LoRA 模型，在此表示感谢！如有侵权，知会立删。
Portrait and Art Photography Optimization Base Model. The model has been specially optimized in composition, light and shadow, oriental face shape to further strengthen the sensitivity and adaptability of the face model.
Compared to the Flux.1 Dev original model, this version is more realistic and richer in terms of ethnicity and face shape, and the example drawings of the face model are from the LoRA model of the following authors, thanks in advance! If there is any infringement, it will be notified and deleted immediately.
https://civarchive.com/user/el_fluppe
https://civarchive.com/user/wolfcatz
https://civarchive.com/user/seanwang1221
https://civarchive.com/user/nawusijia
模型使用简单指南：
基本构图：deis+simple / euler+beta; 噪点更多：ddim/dpm_2/dpmpp_2+beta/beta57/sgm_uniform; 细节更丰富、更具想象力：heunpp2+ddim_uniform; 放大：UltimateSDUpscale/TTP; 胶片效果：添加 lut(35mm/AGAF/Kodak); 或者基于您自己环境的最佳测试组合。steps 20-30。工作流参见例图。
Basic: deis+simple / euler+beta; More noise: ddim/dpm_2/dpmpp_2+beta/beta57/sgm_uniform; More detail, more imaginative: heunpp2+ddim_uniform; Upscaler: UltimateSDUpscale/TTP; Film effects: add LUT (35mm/AGAF/Kodak); Or the best combination of tests based on your own environment. steps 20-30。 The workflow is shown in the example POST image.
Also on Huggingface.co
一个有趣的脸模控制测试示例 (An interesting face model LoRA control sample)：
===========================================================================
DedistilledMixTuned Dev V3.0:
中国农历蛇年钜献! (Great upgrade for Chinese Snake Year!)
V3.0 版模型全面升级，可能是目前 Flux Dev 微调模型中，模型能力最均衡，LoRA 兼容性、真实感、出图质量和艺术创作力最接近 Flux Pro 的模型。(为了方便评测对比，本模型的Seed与原Dev模型基本对齐)
Fully upgraded Version 3.0, it may be the best model in the current Flux Dev fine-tuning models. Have the very good balance in model capabilities, LoRA compatibility, realism, image quality and artistic creativity closest to the Flux Pro model. (For evaluation and comparison, the seeds of this model are basically aligned with the original Dev model)
V3.0 版模型使用指南：
模型通过分层融合技术，去除反蒸馏干扰，与原版 Flux.1 Dev 模型完全兼容，具有更高的 LoRA 权重敏感性。1024x1024 分斌率以下，建议 euler/deis + normal/beta/simple 等，1024 - 2048 大分斌率图像，建议 ddim/dpm_2/dpmpp_2m/heunpp2 + ddim_uniform/beta.
细节最强：dpmpp_2m+beta，艺术性最好：heunpp2+ddim_uniform
建议：KSampler, 20-30步。工作流请参见图：https://civarchive.com/images/53432419
The model is fully compatible with the original Flux.1 Dev. Had removed the de-distillation interference, and has a higher sensitivity to LoRA weights. For 1024x1024 and below, euler/deis + normal/beta/simple, etc., 1024 - 2048 for large binning images, ddim/dpm_2/dpmpp_2m/heunpp2 + ddim_uniform/beta.
More details: dpmpp_2m + beta, More artistry: heunpp2 + ddim_uniform
Recommended: KSampler, steps 20-30. Workflow of the model pls ref: https://civarchive.com/images/53432419
===========================================================================
DedistilledMixTuned Dev V2.0:
2025 新年献礼！经过一个多月打造，在 v1.0 基础上全新升级 V2.0 版本，以照片级的真实感，在细节体现、出图速度、LoRA兼容性、光影和谐度等方面达到一个新的平衡。
2025 New Year Gift! More than a month of training and fine-tuning, The V2.0 version has been upgraded based on v1.0, and has reached a better balance in detail reflection, drawing speed, LoRA compatibility, light and shadow harmony with photorealistic realism.
===========================================================================
DedistilledMixTuned Schnell V1.0:
可能是目前基于 Flux.1 Schnell 调制的各种模型中，快速出图(4-8步)，遵循原版 Flux Schnell 构图风格，提示词还原能力强，且在出图质量、出图细节、回归真实和风格多样化方面取得最佳平衡的开源可商用 Schnell 基础模型。
Only 4 step, The Model may achieve to the best balance in terms of image quality, details, reality, and style diversity compare with other tuned of Flux.1 Schnell. and have a good ability of prompt following, good of the original Flux model style following.
Based on FLUX.1-schnell, Merge of LibreFLUX, finetuned by ComfyUI, Block_Patcher_ComfyUI, ComfyUI_essentials and other tools. Recommended 4-8 steps, usually step 4 is OK. Greatly improved quality and reality compare to other Flux.1 Schnell model.
===========================================================================
DedistilledMixTuned Dev V1.0:
可能是目前快速出图(10步以内)的 Flux 微调模型中，遵循原版 Flux.1 Dev 风格，提示词还原能力强、出图质量最好、出图细节超越 Flux.1 Dev 模型，最接近 Flux.1 Pro 的基础模型。
May be the Best Quality Step 6-10 Model, In some details, it surpasses the Flux.1 Dev model and approaches the Flux.1 Pro model. and have good ability of prompt following, good of the original Flux.1 Dev style following.
Based on Flux-Fusion-V2, Merge of flux-dev-de-distill, finetuned by ComfyUI, Block_Patcher_ComfyUI, ComfyUI_essentials and other tools. Recommended 6-10 steps. Greatly improved quality compared to other Flux.1 model.
GGUF Q8_0 / Q5_1 /Q4_1 量化版本模型文件，经过测试，已同步提供，将不会再提供别的量化版本，如有需要，朋友们可根据下面提示信息，自己下载 fp8 后进行量化。
GGUF Q8_0 / Q5_1 /Q4_1 quantized model file, had tested, and uploaded the same time, over-quantization will lose the advantages of this high-speed and high-precision model, so no other quantization will be provided, you can download the FP8 model file and quantizate it according to the following tips.
===========================================================================
Recommend:
UNET versions (Model only) need Text Encoders and VAE, I recommend use below CLIP and Text Encoder model, will get better prompt guidance:
Text Encoders: https://huggingface.co/silveroxides/CLIP-Collection/blob/main/t5xxl_flan_latest-fp8_e4m3fn.safetensors
VAE: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae
GGUF Version: you need install GGUF model support nodes, https://github.com/city96/ComfyUI-GGUF
Simple workflow: a very simple workflow as below, needn't any other comfy custom nodes(For GGUF version, please use UNET Loader(GGUF) node of city96's):
===========================================================================
洗去蒸馏油腻，回归模型本真，致力打造最纯正的 Flux 优质底模！
Wash away the distillation and return to the original basic.
如果您使用后觉得模型不错，请多多返图，谢谢！
If you feel the model is good for you, please post the image here, thanks a lot!
Thanks for:
https://huggingface.co/black-forest-labs/FLUX.1-dev, A very good open source T2I model. under the FLUX.1 [dev] Non-Commercial License.
https://huggingface.co/black-forest-labs/FLUX.1-schnell, A very good open source T2I model, under the apache-2.0 licence.
https://huggingface.co/Anibaaal, Flux-Fusion is a very good mix and tuned model.
https://huggingface.co/nyanko7, Flux-dev-de-distill is a great experimental project! thanks for the inference.py scripts.
https://huggingface.co/jimmycarter/LibreFLUX, A free, de-distilled FLUX model, is an Apache 2.0 version of FLUX.1-schnell.
https://huggingface.co/MonsterMMORPG, Furkan share a lot of Flux.1 model testing and tuning courses, some special test for the de-distill model.
https://github.com/cubiq/Block_Patcher_ComfyUI, cubiq's Flux blocks patcher sampler let me do a lot of test to know how the Flux.1 block parameter value change the image gerentrating. His ComfyUI_essentials have a FluxBlocksBuster node, let me can adjust the blocks value easy. that is a great work!
https://huggingface.co/twodgirl, Share the model quantization script and the test dataset.
https://huggingface.co/John6666, Share the model convert script and the model collections.
https://github.com/city96/ComfyUI-GGUF, Native support GGUF Quantization Model.
https://github.com/leejet/stable-diffusion.cpp, Provider pure C/C++ GGUF model convert scripts.
Attn: For easy convert to GGUF Q5/Q4, you can use https://github.com/ruSauron/to-gguf-bat script, download it and put to the same directory with sd.exe file, then just pull my fp8.safetensors model file to bat file in exploer, will pop a CMD windows, and follow the menu to conver the one you want.
LICENSE
The weights fall under the FLUX.1 [dev] Non-Commercial License.
Description
FAQ
Comments (102)
Great work introducing PAP!
I wanted to ask if you plan to ever release fp16 versions of your model for download. Thank you!
Thanks a lot for your appreciation! it is so sorry, No, Only fp8.
@wikeeyang may I know why?
@Queezy Too difficult for me to uploading a large than 23G file. transfer fp8 for me need to repeated many times due to network.
@Queezy and for my model, the generate image of fp8 model is enough detail.
Dear wikeeyang!
GGUF (BF16) ~ 22-23 Gb ?
Flux1-DedistilledMixTuned v3.0 (not PAP) BF16 ?
it can be done with large RAM
If the problem is buying a powerful video card, I think collecting donations will help. Many will come to the rescue to get their hands on a 16-bit, 32-bit Flux1-DedistilledMixTuned v3.0 (not PAP)
@alex456mint1 Thanks a lot for your advice and willing to help me! maybe wait for the time is right, now, most people don't really know the true capabilities and values of the model v3.0.
请问为啥模型分类是Flux.1 D Refiner？这个和普通的Flux.1 D分类有什么区别吗？
我下的是最新那个，单独用还好，可是为什么叠加自己的lora，画质就变得极其糟糕，不知道问题在哪里？
不知道，或者你可以分享你的 LoRA 出来我测试一下，我测试了很多 C 站的 LoRA 都没问题。
@wikeeyang C站的lora一样的有问题，不知道是采样器的问题还是什么原因
@youknow8209144529 你是用 ComfyUI 还是 WebUI，C站的LoRA大部分都质量挺好，那肯定是你使用方法或参数方面有问题，如果是 ComfyUI，拉个人家带工作流的例图，看看人家的参数学习一下就都会了。
can you convert this to nf4?
Sorry, may you try Q4_1? I had published it
or you can use the tool convert it by yourself according to the instruction in the model introdction.
@wikeeyang Okay, I'm using WebUI Forge because, as far as I know, I can only run Flux NF4 on my device. I'll try Q4 then! :)
@patrik WebUI Forge support GGUF, I think it is no problem for you to use Q4_1, and as I know God Min (lllyasviel Chinese name is Zhang lv Min, we called him MinShen 敏神 😂) was stop to update the project long time, I suggest you tune to use ComfyUI, it is more flexible and scalable.
you can convert it to NF4 by yourself in Forge
@OliviaRossi How?
@NoobFromEgypt in the tab, where we used to merge models - now there is a dedicated section on the very top of it - select the checkpoint, vae, clip, text-encoder and NF4 or NP8 and start the process - it will take around 10 min - maybe more
@OliviaRossi thank you
@NoobFromEgypt u r welcome!
@OliviaRossi Much thanks for your kind suggestions!
@wikeeyang <3
Hi need help. i applied the same setting but doesn't work on Comfyui.
this is your setting Steps: 25, CFG scale: 1, Sampler: deis, width: 832, height: 1216.
Why is this?
I don't know what happend for you? any error message? or other model normal run but only my model don't work, if you need some help, pls provide more message detail, thanks.
@wikeeyang Ok it turns out i didn't install the missing nodes. i did and now it works. Thank you for replying. Took me 2.3 minutes per image with 832x1280 on my RTX 4070 but after 30 minutes of conitinous generation, it took me 68 seconds per image. It's quite painful at first but later its getting a bit faster.
It's a great model. I'm comparing head to head with The Amaranta Experiment checkpoint.
Thank you.
@haidensd58757 OK, Good news, Any problem or suggeststion let me know feel free.
大神，反馈一下：V3.0-PAP对LORA会有点污染，V3.0表现非常好！
谢谢您的反馈！主要体现在哪方面的污染？从底模来说，3.0应该比3.0-PAP更纯净，PAP毕竟做了东方人脸型方面的适配，训练素材多了、杂了一些。方便的话请描述一下，我看能不能改进，谢谢！
@wikeeyang 主要是人物lora的脸部影响比较大，出来的图片人物面部不是很像了。不改变任何参数情况下更换V3.0就没有影响了。
@wzb2000wcy388 这就有点怪了，有没可能只是某个特定的 LoRA ？你看我的例子和 POST 样图，我测了好多人脸 LoRA，都是 v3.0-PAP 比 v3.0 更敏感，人脸相似度更高。或者你方便的话，把你测试的 LoRA 名称或链接给我一下，我来测一下，谢谢！
any updates for schnell unet ?
Sorry! No plan, due the textures and details of the de-distilled Schnell model really not enough. 🤣
Help!
Hey, which is better, the FP8 or the Q8?
As far as I know, ggufhase better results, but here fp8hase more downloads and reactions than the gguf one
Haha, personally, I don't think it's much different between fp8 and Q8_0, but it is rumored that Q8 is a little bit better, I think it just depends on your environment and usage habits.
@wikeeyang Got it captain!
Actually, I have only used NF4 and GGUF models, so I was confused about FP8.
@wikeeyang i also thoughtI used to think GGUF was better than FP8, but when I noticed that FP8 had more downloads and the images were really impressive, I got a bit confused!
generally speaking Q8 is closer to fp16 but is a bit slower of fp8
@3dave @ALFARANKO As I test, the fastest way of quantize method is SVDQuant now. but this quantization process is very slow, usually require a high graphics card like H100 or H800 and a few hours to more than a dozen hours, therefore, most people are not capable of doing this thing.
how can i remove flux chin? as i can see some of your works don't have flux chi.
your need use many no chin face image(for example chinese face model) to train and fine the Flux.
Kudos to @wikeeyang for this excellent model! This is by far the best FLUX based model I have ever used, and I am not exaggerating. I have been working almost exclusively with v3.0 Q5_1 (non PAP) and it has handled everything I threw at it extremely well! I have generated photorealistic, comic style, anime style, watercolor style, and I have yet to find a weakness for this model!
It also plays very well with all the LORAs I tried alongside it. I am more than happy with the results this model produces.
Wishing you all the success you deserve! Thank you!
Thank you very much for your detail testing, feedback and appreciation.
what loras are recommended to make really good realistic images?
Sorry, there is no recommendation really, due to the quality of LoRA is uneven, depending on the quality of each trainer's image and tag materials, environment, requirements, etc., then only be test and tune the parameter more times.
Can I ask, if its possible to use fluxgym to make a lora with ur model? im kinda new to this.
What u recommend using to UPSCALE? Or should im trying to use SDUltimate upscale, but nothing suits well
@TropicalCat Do you just want to ask something, or do you want to generate image with a model? If you want to learn how to use Flux model to gererate a good image, pls study the image in Civitai of other people's who posted, it include all parameter how to generated it and Upscale it.
@wikeeyang Ive made a lora for a charater (flux) And really liked the results with ur model, but idk how to make the nsfw haha
@TropicalCat Haha! I don't pay attention to NSFW, but this aspect should actually be a drawing skill things, you can use my model to generate a image you like, and then use my friend's fill(NSFW) model to redraw and undress. model link: https://civitai.com/models/958009?modelVersionId=1445047, Ha ha...
@wikeeyang Thanks!
pap版本非常喜欢，但人脸lora出图不太像，看了一下评论似乎别人也碰到类似问题，lora是基于原版dev炼的，同一个lora，3.0就好不少。另外想问问是不是对全身照不太友好？经常出现腿部崩坏，要么就人身比例不协调，希望老师有时间可以针对做做优化？
您好！感谢反馈，有测试例图吗？可以上传给我看看。或者，您可以参考我POST的例图，里面有工作流，我自己测试的结果，PAP 版本对 LoRA 脸型还原更好，您看例图里面的明星像吗？我也测试了一些身体姿势，感觉基本还好，大部分姿势要求能出来，这块我会继续留意，并有可能的话，继续优化，谢谢！
当然，3.0 模型跟 Dev 的遵循性应该是更好一些，这个不用怀疑。另外，LoRA 由于训练素材质量、训练者的个人审美、训练环境等因素，干扰项比较多，所以不同 LoRA 效果偏差较大一点，如有哪些 LoRA 效果相差较大，可以提供一下链接给我参考，我可以在下次调教时加入测试和修正。毕竟我时间有限，测试的 LoRA 和图片数量都比较有限，谢谢！
@wikeeyang 非常感谢老师的回复！关于LoRA的问题，其实我更倾向于是自己炼的问题，因为是第一次炼人脸没什么经验，训练图质量本身不够高，也有过拟合的可能，您的例图我看了，效果是非常不错的，所以倾向是自己的问题哈哈哈。（lora例图恐怕不太方便提供，望体谅！）
关于姿势方面，姿势以及构图是ok的，提示词遵循度也很好很稳定，只是在抽全身照的时候，会偶尔的出现人身比例不协调的情况，例如脚很短，或者很瘦小的腿部之类的情况，对比半身照来说稍显不稳定，这一点我也不知道是什么原因了。
最后，再次感谢老师您出色的工作，3.0和PAP目前都在大量使用，非常优质的模型，个人非常喜欢！
@hihi94 谢谢您的反馈与建议！训练一个能用的 LoRA 门槛好像不太高，但要训练一个高质量的 LoRA，对训练素材和 TAG 打标和训练参数的调教还是有很高要求的，这方面需要很多次的经验积累，祝你越训练越成功。关于您提到的 whole body 人像肢体协调性抽卡的问题，这个我大概能理解了，这应该是 PAP 模型镜头调教的问题，因为模型主要聚焦近焦的人脸特写和半身的艺术摄影或肖像为主，特别强化了近距离的皮肤细节，因此在全身摄像和景观方面的聚焦、构图有所减弱，如果想出全身照片和景观类的图，我建议使用 v3.0 底模，因为全身人像其实对皮肤细节要求就不高了，v3.0 模型在全景和大焦距方面的构图会好很多。谢谢！
非常好用的模型,找了无数Flux模型都不可避免的油腻,只有你这个完美真实, 看来我可以删除其他模型了哈!不知道是否会更新schnell 4步的Q4? 我发现任何同模型用两个Q4, 比如总步数10,先用schnell 跑前4步,在接正常(可带加速Lora)跑后面几步,最终的效果出奇的好, 之所以2个Q4主要我16G显存占用不会太高,和单独FP8差不多,而且复杂工作流也跑,前4步抽卡也快! 是否有qq 微信群啥的
非常感谢您对本模型的测试、喜欢和肯定，v3.0 版模型，毫无疑问是目前所有 Flux 开源微调模型中，唯一种子(Seed)能与官方 Dev 模型基本对齐，底模质地最纯净，模型的材质提升和出图能力最强的模型。由于本人喜欢埋头研究技术，没时间打理QQ或微信群。有任何问题或建议，欢迎留言，留言必复，谢谢！
至于 Schnell 4 步模型，暂时没找到进一步提高材质和细节的方法，目前相比 Dev 模型，在材质和细节方面没法突破，另外，暂时也没法克服步数太小，对原来基于 Dev 模型训练的 LoRA 的兼容性问题，或许迟点找到方法了，会第一时间更新。谢谢！
Thank you so much! wikeeyang, this is an awesome underrated model! I've tried many different Flux models, but this one is special. I did it at 1024 x 1344 resolution with no upscaling. I've only tested fp8. And I can't stop, it's so good.)) DedistilledMixTuned Dev V3.0 (not PAP yet). DedistilledMixTuned Dev V3.0 is the best modified Flux-dev model. Are you not planning a DedistilledMixTuned Dev V3.0 fp16 or fp32? That would be a sensation!
Thank you very much for the likes and appreciation! based on my evaluation of hundreds of fine-tuning models from Flux Dev open source, my v3.0 model is the only one that can be basically aligned with the official Dev model on the seed, and the texture of the this mold is very pure, and the details, materials and generation ability of the model are greatly improved compared to the official Dev original. I think the FP8 is enough for daily use now, and the higher accuracy may be shared when I study to a certain extent, thank you again.
在简单的背景下，出图很容易出噪点
有这问题，可以加背景描述提示词来消除或缓解，模型整体噪点加强是为了去除 Dev 模型皮肤的塑料感和油腻感。
Hi its possible to use FluxGym to train with your model and use as lora? (If yes how can I do it?) haha tysm good ckpt!!!
Should be ok, just use it as normal offical Dev model.
@wikeeyang Can you give the training parameters with which to train on a person?
@Queezy This is not easy to give, the basic parameters are given by the flux trainer, and the specific parameters need to be adjusted according to your own training environment and training data set.
I understand correctly, this retains the flexibility of original dev (no added woman bias etc.) and has CFG, so I can use negatives, and is still compatible with regular flux dev loras?
Pls use it like the official Flux.1 Dev model, the LoRA is fully compatible, but negative prompts seem to have not effect.
bro, may i post this on tensor art so i can use this model? thank you so much!
Hey bro, I allow you can forward this model, please follow the spirit of Open Source and indicate the source of this model, Thanks.
version 3 is no longer de-distilled?
So CFG (negatives) will no longer work?
All my model isn't pure de-distilled, needn't true CFG parameter, just use it as normal Flux.1 Dev model, I only mixed de-distilled model of Dev some detail and metrial layer before finetune.
@wikeeyang ah, so it's trained over flux1 dev but you added a little bit of de-distilled into it after training?
@ForeverNecessary737716 before finetune training.
想要NSFW Q///Q
Pls ref to my friends https://civitai.com/models/958009?modelVersionId=1550408, is based on my model and support NSFW. 🤣🤣🤣
try to develop your schnell more please.
Thanks! yeah, I just to try and find some method to inprove the detail and metrial of Shcnell.
@wikeeyang your model one of the best if it's not the best .. as schnell 'i tried few' , will be good training your schnell more and focus to improve it more , thanks
@amazingbeauty Thanks a lot! I will try it.
Thank you for this nice model! There's just one issue that I found. I'm getting very pronounced splotchy noise in the backgrounds. This is especially noticeable when the backgrounds are blurred. It's as if I'm using Detail Daemon with heavily increased "detail_amount" parameter. If anyone else has this same problem, here's how I managed to solved it to a large extent. I'm using ComfyUI by the way.
Download the "Detail Daemon" node, connect it after the "KSamplerSelect" node, and then connect the "Detail Daemon Sampler" node to your Ksampler. You will need to use a Ksampler that allows the usage of a separate sampler node such as "SamplerCustomAdvanced". Then, use these settings in the "Detail Daemon Sampler" node:
detail_amount: -1.00
start: 0.72
end: 1.00
bias: 1.00
exponent: 1.00
start_offset: 0.00
end_offset: -1.00
fade: 0.00
smooth: true
cfg_scale_override: 0.0
You may need to tweak these settings a bit, if you find they don't work too well in your particular case. Mostly the "detail_amount" parameter, the "start" parameter, and the "end_offset" parameter. The lower you set the "start" parameter, the more noise will be removed, but you also may start losing details in your images. I hope this helps somebody. :)
Thanks a lot and appreciate your sharing of experience in use this model, so that may let more people can use this model better, thanks again.
Can you share workflow or png with it?
DedistilledMixTuned Dev V3.0 (not PAP) - best Flux.1 class model in my opinion and superior to Flux.1-dev, adds a lot of skillful or cute details. I use 40 steps with dpmpp_2m + beta at 1024 by 1344 resolution and 25-30 steps for heunpp2 + ddim_uniform. The results are often beautiful, as far as the 8-bit model goes. Unfortunately 16-bit and 32-bit cannot be tested yet.
The model allows mild non-violent erotica, i.e. not all of NSFW, but only part of it.
I previously liked this model Freepik_flux.1-lite-8B with some disabled blocks in the modification process.
https://huggingface.co/Freepik/flux.1-lite-8B (15-16 Gb)
But wikeeyang's DedistilledMixTuned Dev V3.0 (not PAP) is superior to Freepik_flux.1- lite-8B
DedistilledMixTuned Dev V3.0 (non PAP) Q8_0 is a more accurate model (key computations are still performed at 16-bit precision, more steps can be used) than FP8
DedistilledMixTuned Dev V3.0 (non PAP) FP8 is a more artistic model (key computations are performed at 8-bit precision) than Q8_0, in subjective opinion.
Both models are beyond competition in terms of artistry and colorfulness, painterliness, which are revealed by skillful Prompt (with adjustment of the weight of the whole Prompt from 0.7 to 1.5 or with fine balancing of logical parts of the Prompt separated by a point with the same weights, the first part of the Prompt with focus and accents, descriptive part1, descriptive part2, additional description, detailed style, color contrast + color saturation + textures, light, tech. lens characteristics and quality, accents), using t5_xxl FLAN (16- or 32-bit) and large Clip_l (16- or 32-bit, with only TE or full versions).
Thanks a lot for your detailed test and professional evaluation.
Which version will work on 12Gb GPU, with reasonable inference time, and capable to generate NSFW females?
Asking because downloading single 11Gb file will take several hours for me
My model can't generate NSFW iamge directly, you need plus some NSFW LoRA at least, if you use ComfyUI, I think 12GB VRAM should be able to be use fp8 or Q8_0 reluctantly by CPU offload, Q4_1 or Q5_1 should normally.
Thanks for your work this is a very nice model
Thanks a lot. 🤝🤝🤝
can anyone tell me how long it will take me to generate 1 image in a 12 gb VRAM ( rtx 3060 ) with this model ?
It is depends on the size of the images you generate, for example 512x512 or 1536x1536, and some other parameter for example Schduler + Sampler, I think may take 0.5 - 2 minutes.
huge models! thanks so much.which one u prefer for best result? i mean which one is the best one?
Thanks a lot, the best one should be the next...🤣, V3 should be ...
@wikeeyang is this model something like flux sigma vision? i can use my own lora models on this but they dont work on sigma vision model
@ehsanbbaxe Should be a little like, I just append some de distilled weight to tuned Flux.1 Dev model, then you can use my model as Dev model, but sigma vision was more like de distilled model.