yeah bbox works quite well with minimal prompt leakage

@Lynx2025 Model is insane, my Krea prompts are simplistisc AF (I don't use LLM's) but this model understands basically the most basic stuff and it gets right, it has some deformation with fingers and sometimes it gets body horror's but I believe that is a issue with the turbo LoRA, but with time it will be a great model.

Krea2 still good for the near future. 12 buzz/image with Turbo. A lot of LoRAs and checkpoints. It'll take a few months for MOSS to truly unseat it.

comfyui works fine. update comfy to latest build instead of stable, or just wait a bit for a new stable update

I don't think it's superior to Krea2 honestly, but definitely a nice mix of Flux and Anima. If they release MOSS XL or something like a 12B parameter model, then yeah possibly. The license is very good as well.

That was quick.

You legend

Adding on for anyone interested in training as well, the paper only covers their dataset preparation methodology briefly.

For a better idea of how the MOSS team captioned their images, do check out this guide under their GitHub repo -> https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md#post-training

there is a turbo lora
https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd

I used for the following images:
https://civitai.red/images/145922389
https://civitai.red/images/145919628
https://civitai.red/images/145915725
https://civitai.red/images/145899953

@aj_prime8517 man that was fast :D thank you very much

@aj_prime8517 Super fast gens with the turbo lora thx. 1.5mp takes 10 secs now lmao

@mu5ic3oy337 yep. 12second now for me and quality is great too.

@infernahermit846 you are welcome!!
But all the praise for the creator of the lora, I am only an user of it

@aj_prime8517 I understand but I had no idea there was a lora before you mentioned it.

quality above speed, Turbo loras kill quality, the model is fast to begin with if you play with the settings and have a capable GPU

@hRUSTIN474 Every model is fast if you have a capable gpu, thats not the point and we dont have capable gpus. also the lora with 15 to 25 steps generates superb images with high quality and since we are using 1 CFG it's fast.

@infernahermit846 other option is to use the lora with 8 steps and CFG 3 that is I use with my images, in generation time should be equivalent to 16 steps with CFG 1, I don't know which combination provide better quality (concept that on the other hand is subjective) and maybe it will depend on the kind of images are generated, for instance I don't usually generate realistic images of people

@aj_prime8517 True. for me any CFG higher than 1 burns the image a little bit but at 15 to 25 steps gives me a really high quality output. it's fine with 8 step 1cfg but body hair, grass becomes plastic(ish) and skin gets too soft.

But for cartoon/3d looking stuff cfg3 works really well, like the bee you made.

All the images I uploaded here were generated in 8-steps with the unfinished Turbo Lora from Lodestones. CFG1.0 Euler-SGMUniform. Seems to work pretty well.

@Talanca Yep, me too, uploaded 8 images, with the range from 10 to 25 steps.

Same people are focusing on an anime version. Imagine that.

No innate support for artist tags is a huge negative sadly ;_;

@DaddysLilMistake Yeah, but it's always better to use a lora for that as it will always give a better result
One thing i like a lot is that it's able to get characters very well
I even tested on some obscure characters and the model was able to make them

@DaddysLilMistake Had a feeling "kills Anima" was a bit overstated.

Am I missing something? I've done a lot of tests, and at it is right now (only for anime), Anima is MILES better in overall quality. Strange, becuse I could say everything you just said, but swapping the two models. Animadoesn't need expensive prompting at all for me. However, I am super excited, because the model is still still new and they are going to release an anime version of it.

Could you tell me how to use this?

https://github.com/OpenMOSS/MOSS-Image/tree/main/comfyui
Official node pack doesnt have a custom scheduler node

I'm not a comfy user but according to Grok:


1. Create a folder for the node

Bash

cd /path/to/ComfyUI/custom_nodes
mkdir my_moss_scheduler   # or any name you like

2. Put the .py file inside

Copy your ready Python file into the new folder, for example:

text

custom_nodes/my_moss_scheduler/moss_scheduler.py

3. Create init.py

In the same folder create a file named exactly init.py with content like this (adjust the import and class name to match your file):

Python

from .moss_scheduler import YourSchedulerClassName   # change to the actual class name inside the .py

NODE_CLASS_MAPPINGS = {
    "MOSS Custom Scheduler": YourSchedulerClassName,   # the key is the internal name
}

# Optional – nicer name that appears in the menu
NODE_DISPLAY_NAME_MAPPINGS = {
    "MOSS Custom Scheduler": "MOSS Custom Scheduler",
}

__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]

How to find the class name: Open the .py file and look for a class that has:

INPUT_TYPES (or the newer define_schema)

RETURN_TYPES (often ("SIGMAS",))

FUNCTION = "get_sigmas" or similar

CATEGORY = "sampling/custom_sampling/schedulers" (or similar)

Use that class name in the import.

4. Restart ComfyUI completely

Close the server and start it again. Refresh the browser (Ctrl+F5).

After restart, search for the node name in the node menu (right-click → Add Node, or the search bar). It should appear under the category defined in the file (usually sampling/custom_sampling/schedulers).

5. How to use it

Most custom scheduler nodes output a SIGMAS tensor. Typical wiring:

Connect your MODEL (and steps / denoise if the node asks for them) into the custom scheduler node.

Connect the SIGMAS output into a SamplerCustom, SamplerCustomAdvanced, or KSamplerAdvanced (the sigmas input).

Do not use the normal scheduler dropdown on a regular KSampler when you feed custom sigmas.

Quick checklist if it doesn’t appear

Look at the ComfyUI console / terminal for import errors (red text).

Make sure the folder contains both the .py file and init.py.

Class name and import must match exactly (case-sensitive).

No syntax errors in the Python file.

If the file already contains NODE_CLASS_MAPPINGS at the bottom, you can sometimes just drop the single .py into custom_nodes/, but the folder + init.py method is more reliable

Are you sure this qwen was not finetuned?

@Shio_N it really just seems like someone at civitai messed up. Like whatever internal reference they are using to the file is different from that public facing reference. It's definitely not right, but I think this is equivalent to "the intern fucked up". You could also just go download the weights from Comfy themselves.

no

I'm using my Intel Arc b580.

(Intel cards) For fast enough int8 you should rewrite default pytorch int8 processing - asked AI to do it. Currently 1.5 seconds per step (was 4 seconds per step in int8 initially), but can be better as bf16 is initially 1.5 seconds per step without any improvements - it just have better software support. I have used only free coding agents for it. Waiting for limits lift to continue optimization.

no sir... 9070XT here (2Mp latent) : "100%|██████████| 50/50 [01:20<00:00, 1.61s/it]"

;)

I'm running it just fine without NIVDIA

That’s the same for me, bro lol

I cleared out the storage for Krea 2 and Qwen 2.1, but Krea 3 is already coming out...

remove ALL old image models

@qqq_qqq_qqqq807 I find Krea2 a bit better for now in the LoRA space, and Illu is the absoulte king of Loras, so much niche stuff I can't really find it, now Qwen got nuked, H3 is still there but I don't have many LoRAs for it, just need more storage by now

https://www.reddit.com/r/StableDiffusion/comments/1x7zf1r/hey_guys_im_one_of_the_core_developers_of/

You found a lot of information on this topic, awesome! I've subscribed, keep it up, bro :)🔥

thanks for your information！

Better than krea2 I think. No need for uncensor loras, has no plastic skin or vae issues and is faster.

OH YES! It is completely uncensored. It can do tentacles as well! :D :D

@bpbp2 my first impression is its on par. it feels less 'alive' than krea2 though, just a feeling i get after looking at a bunch of images. seed variety is much better though.

It does, i believe it doesn't have artist knowledge
I even tried some obscure characters and it did well

@danilen52 nah model doesnt know a donghua characters, i did generate 3 last images with popular ones - yaoxixue, 姚惜雪, 仙逆 and feng nianyun, 百炼成神 封念云

Maybe! But now our priority is to release an anime version of the model.
yeah, anime data accounted for a tiny fraction, and we're not really satisfied with its current performance.

you can get it to Krea2 with doubling the image size and step count. Like around 3-4MP it is kind of insane. It will be slightly rough sometimes, but when it works it's indistinguishable. Well that's not true, this looks better because of how human the people look. However Krea2 is still better at very technical anatomy or with more than two people. I honestly think the low parameter size is what helps it be so realistic, it is able to combine many different concepts when you ask for one, so the output looks more like you'd expect. It's a balance game, I think. This low parameter count does mean we miss out on more accurate copy-righted material though, which Im sure people will dislike in the anime/cartoon space.

krea 2 worth nothing... Even SD1.5 is more creative than krea. Not sure why people want to generate the same images with slightly different content. Each new result feels like previous and better anatomy/quality don't help when it's initially no what I want to see. Even here... "Base" is much better than "SFT" in terms of creativity, but people still use that horrible finetune that generates very similar content.

@Shio_N Krea2 you can manipulate noise schedules to increase seed variance. But models like qwen or zit or Klein really struggle with it. Chroma is the only other model that has gotten close to this (MOSS)

I just used my Krea2 flow - switch the clip to moss-image and switch the models out.

You do need to update your comfy to the latest version.

@Fatbuns did it already, i get '
The value moss_image for Load CLIP's type is not available.'

I'm using latest protable version. it is not yet added to comfy's templates either. but works for me.

@yoloswagg45 Not sure if it will help (or if you already tried this), but I was getting the same error after installing the latest requirements, did a git pull in the comfyui folder and the moss image option for the clip appeared after that.

you just gotta update to nightly I think. It was a little weird for me at first too, but using the update scripts in the repo fixed it

yep same here, tried the official workflow and some others and every time end up with errors

@makiaeveli could you plz for the love of God explain, which script you mean? i've tried everything, but nothing seems to work for me

@Bergamoth like in the files on your computer. "ComfyUI_portable/update/update_comfyUI.bat". IF you don't have windows you can just open the script and run the python script in that same folder with arguments presented in that bat file. I think when it didn't work this time, I ran the "stable" script, then the just "update" script. Make sure to close the windows and everything. Sometimes restarting the computer helps if for some reason it doesn't still work. But if it is still "broken" after those two things, you could try forcing it with the "git pull" command, but you should ask google real quick to get the branch right.

@discordia100956 Umm which file did you pull? and where did you put the folder, im kinda new with these stuff..

@harga5643757 no, he means "git pull". that is effectively all the /update/update_comfyui.bat script is doing (it also makes a backup etc)

@makiaeveli Can you do a git pull with the official desktop version? I thought this was only possible on the portable version. Ppl often talk about comfyui like there is just one version.

@harga5643757 honestly I am waiting a couple of days until comfy desktop gets an update

@Yahoo384 I have no idea i just use the portable version. if you can select the nightly version in whatever capacity, it should be there.

@Yahoo384 I'm not using the portable version. As makiaeveli said it is basically an update all, it pulls the latest commit from the comfyui github. In your terminal start the venv you use for comfy (if you are using one), make sure you are in your main comfyui directory inside the terminal for example home/comfyui and just type "git pull" and hit enter, it will update comfy to the latest version.

@discordia100956 the latest desktop version 0.42.1 has added support for MOSS, it is working now.

Yes, but it's extremely sensitive concerning correct hand and feet anatomy, and hardcore NSFW stuff depends a lot on the seed and you need some luck.

Yeah it seems crazy but it legit works, and as @Experimental_dAIver stated, hardcore stuff is a dice roll, I tried cunnilingus (as that's the hardest thing for ai to do nsfw) and I get it like 1/5 seed attempts.

For 5B checkpoint it's even great, can be compared with ZIT/ZIB with more parameters. Don't ask about FLUX Klein, 2 and 3. Just forget them.

It has various minuses in textures but TBH, I don't look at them because the subj is a loyal to weak GPUs with good perfomance. SDXL/SD1.5 were also poor someday.

use 25 to 30 steps, 50 is unnecessary

Thanks, can I ask you how long does image generation take compared to 8 step Krea 2?

it's kinda weird with art in general, or at least I'm very bad at prompting for art. it mixes styles a lot, or is good at it. I can't ever get consistent "cartoon" or "anime" photos at all. I have seen them though, just takes good prompts

Tell me your Comfy Backend, Comfy Kitchen and Torch versions? int8 required them all updated to the latest versions and Torch built for CUDA 13

@leonviz Did you update your ComfyUI?

i'm getting good images with sft at 30 steps. i just have to give the model a 2mp latent to work with... ymmv

@tedbiv Thanks for this idea. Had to use my custom-edited version of the "Resize seed from" feature in Forge Neo (the original version has been broken for 4D models) to do the same idea, but that overall improved generation for me at lower output resolutions, along with a slight cfg bump.

@s055599632 i'm using comfy. haven't updated forge yet. running cfg=5, sampler=er_sde, scheduler=simple, steps=30, 2mp latent. using my anima and krea2 prompts i'm getting some killer images.

@s055599632 also found 2mp latent let's krea2 and zit do better with full body images

@tedbiv that's some voodoo mojo lol, never knew bigger latent can actually improve things 

@Bergamoth it gives the models a larger area to be creative. it helps with finer details. like eye details in full body images.

Per their prompting guide, the recommended "worst quality, bad anatomy, watermark, jpeg artifact" negative tags (you can leave out or put in the nsfw in first slot as needed) and putting CFG up past 3 if you have a short prompt are pretty important. I got pretty awful results until I added those on SFT, base is still a horror show with the same level of detail in the prompt, I'd probably have to put in a lot more specificity to get the same out of base.

try use int8, I am 3090 too,1024 res is about 0.5s/it

@calvantsang0808 Wow, that's really fast..

And here I am still tweaking my Flux 1 D with 80+s/it with 3080. I can't move out of my comfort zone.

@egraaha I feel you. Honestly, so far, nothing beats Krea2 Turbo merged with SNOFS & It's realistic lora for me. Speed & quality,so far, I am happy. 8-9 seconds per 2MP high quality image on INT8.

Moss feels half-baked yet.

There is certainly something wrong. I have 3080 and yet the speed is 4.5s/it on BF16 and 2MP resolution.

@munchkin Hm. Is there any chance you could share your workflow, please?

@everylight I used the standard workflow that is in templates of ComfyUI.

Can you share the link to that anime to real LoRA?

Inside Lora models, this site is called "Anime to Realistic Edit".

it's fundamentally your prompt style and negatives. Try out some common negatives like "blurry, out of focus," and make sure your CFG is at least 2, I've had more luck at 3 or 4 with some prompts. If it still blurs, you need to make sure your prompt isn't accidently triggering these things with improper uses of "movement", "focus", "central", etc. That blur you keep seeing is actually fundamentally a strength of the model, it has good understanding of direction and perspective if you find the right words

@makiaeveli What is the way to obtain an image without background blur? The blur even appears on parts of the subject's own body. Could this be a mechanism from the original MOSS model—designed to emphasize text over images—that is affecting the generation process?

My library of prompts, which produces correct images in Krea2, Z-Image Turbo, Chroma, and Flux using the exact same prompts, does not exhibit this issue. I can copy&paste the same basic prompts across different models, and they don’t deliberately emphasize or blur elements in the image.

@Talanca the older models didn't have as good of control over movement, thats what im saying. most of my prompts start with "raw, unedited, crisp amateur photograph of", and most of my negatives starts with " blurry, out of focus, low resolution, low quality, jpeg artifacts, oversaturated, overexposed, underexposed,". If you're using CFG 1, you won't get the negatives, so that could be a problem. I am also using all of danrisi's loras with a low strength of just 0.2, which stabilizes prompts as well. A bunch of my day 1 outputs had blur too, but adding the CFG/negatives and the loras, I only get blurs on movement. If you mention movement at all, expect blur unless you prompt it away.

looking through my photos, I can see like half of them half background bokeh blur and the others don't. It's really about your prompts. like if i wanted to get rid of these blurs i'd say like "static camera, high depth of field, f/22". If you include "DSLR" that can help, but might get more magazine like. I rarely use "cinematic" that will add so much processing. "background blur" could refer to either depth of field or movement. I kinda was expecting you have movement blur, but it very well could be depth of field. This model really likes adding depth of field because the default action is to try and "focus" on what you've prompted. That's why people report heads and stuff getting cut out. It's really wants to find that focal point.

it's not about "text placement", but you're almost there. It's about having better tags/training prompts. It's literally just applying styles to your photos that used to be associated differently. This model just has more control you need to prompt for.

Which models do you prefer for i2i?

Moss doesn't support I2I.

Yep. cirka 4,5s. Upped to 10 to fix some quality issues. Time difference was marginal

Better just wait, one of their devs told on Reddit that they plan to release anime-aimed model with different learning base (so better character recognition, styles recognition etc)

Based on AMA on reddit, the anime finetune is their next release.

They have to meet or exceed WAI. Flat anime faces with line mouths just won't cut it.

@Bergamoth Source? I couldn't find the reddit thread

@hestia it is there https://www.reddit.com/r/StableDiffusion/comments/1x7zf1r/comment/ph93y7y/

@GM_MD WAI is hardly a good model to compare it to. Merge of models and whatnot that completely obliterates diversity of styles.

@munchkin i already didnt even want to bring this up XD

It's the flux vae :P

@Fatbuns ah damn... makes sense then!

Stone and grass, it is what it is, grass is to 'perfect' as well, people need to get more experienced with LoRA's to counter it creating these 'snapshot' LoRA's.

@Fatbuns the f2vae is literally the best open source vae there is

try being more descriptive on the prompt

Are you using ComfyUI and have prompt enhancer in your workflow? If so, disable it. It can mess up your prompt if it contains anything lewd.

try this prompt - best quality, highly detailed, score_9, realistic, photorealism, humanization,\n\nwide angle image capturing the ultimate culmination of fantasy!, tropical beach, huge waves breaking on the beach. in an erotic embrace of the full moon energy,\n\na naked man and naked woman having reverse cowgirl sex by a fire.\n\nthe man is lying flat on his back, the woman is squatting over the mans hips. his erect penis penetrates deeply into her vulva. he is flat on his back in the sand. the man's shiny wet erect penis penetrates the woman's shiny wet vulva. the man's hands are grabbing the woman's ankles.\n\nman, 20 years old yrs, athletic posture, Height: 6 ft. 2 in, Hair: tousled short dark brown hair, a well-groomed short dark brown beard, mustache outlining his jawline, Eyes: Deep blue eyes, dark brown eyebrows, masculine brow, Skin: Sunkissed weathered skin, natural skin texture, broad shoulders, large detailed realistic erect penis.\n\nwoman, 18 years old yrs, long slender legs, Height: 5 ft. 8 in, Hair: Hip-length long straight sun-bleached platinum blonde hair, sleek + windswept, golden highlights, sweeping long platinum blonde tresses cascading past the waist, eyes are complemented by a sharp eye definition, dark eyeliner, + long eyelashes under high-set, neatly groomed blonde eyebrows, sultry gaze, sculpted facial features, youthful heart-shaped facial structure, smooth soft chin, refined small nose, high cheekbones, lips are full with a plush, neutral pout, Skin: Radiant tanned skin with perky tanned medium-small breasts with brown nipples, subtle specular skin sheen, realistic subsurface scattering, soft porcelain finish,\n\nHer long intricately braided blonde hair spills artfully over her shoulders. her eyes are closed with long eyelashes. squatting on the man's erect penis. her legs are spread toward the camera, her body is richly oiled. Her body is leaning forward. her hands are on the man's knees. her mouth is open. expression on her face of blissful happiness overwhelming desire. extreme sensual pleasure. Dramatic chiaroscuro lighting, hyper-realistic texture.\n\nher body is fully upright, squatting on the man's erect penis on a sandy beach at night. her feet are beside the man's hips. her legs spread wide towards the camera. The giant full moon hangs over them. Warm bonfire light illuminates their connection, highlighting the serene expression of sexual arousal on her face. Subtle mist rises from the sand, adding to the idyllic atmosphere,\n\nThe lighting is everything here: intense, warm dark red golden glow from a bonfire floods the scene, creating breathtaking specular highlights on their skin and hair. allowing their silhouette against the huge full moon. emphasizing her incredible curves and long legs. The background features rough water intense ocean setting, water reflecting the massive silver moon with an extremely large full moon. huge waves breaking on the beach, The focus is sharp, highlighting every curve, while the depth of field softens beautifully into an atmospheric, cinematic dreamscape!

I would not say it has that much aesthetic limitations on its own. SFT model has an issue of trading aesthetic flexibility for stability (Mostly it comes to it being naturally predispositioned to realism), yes, so if you seek a more flexible styling, it's better to use base model.

The model itself is 5 GB, plus the encoder (LLM) at 9 GB, and the decoder at 0.3 GB—totaling around 14 GB.

It looks like a graphics card with 16 GB of VRAM will be required.

runs fine on my 3060ti with 8gb vram.

@Corgblam how much time does it take for image gen on your 3060 8vram ?

@aragamin_zero I can run it on my laptop with 8GB VRAM and 48 GB of DDR4. It's kind slow but works

@natasha_26 thats leaking on cpu usage. which is why I asked the "vram" not ram. if it went to pc ram, it will reduce speed a lot and made it unviable for lora training for low vram gpu and many others issues. @Corgblam ofc it will, the real question is will it leak to cpu side of RAM, if possible that needs to be avoided because it will hinder many trainers training LORA capability.

@pedx78207 about 4 minutes for the standard 50 step, but Im using Moss_8step lora and it cuts it down to about 30 seconds.

@Madafada1991 Not sure. Im running it on Forge Neo and have it set to prioritize GPU.

@Corgblam if its 4 minutes, it went to your system ram, if its 30 second then its within your vram.

@Madafada1991 standard Moss is 50 steps, but this lora cuts it down to 8 so its much more manageable.

Much smaller model than Krea2 and a bigger model than Anima. Obviously not that good at anime as Anima, since it was not trained only with anime in mind (it does know some things), but they said they will release anime model later. As for Krea2, it is hard to compare them in terms of capabilities, but I think Moss Image is obviously more uncensored in comparison (even though it says in report that there was filtering in some dataset). Krea2 is also more stylized in my opinion.

it's fun.

Uncensored, better potential than Anima, and bounding boxes can reinforce prompt adherence

请问为什么我生的图里的人物都歪瓜裂枣的，我用base，而且还会发糊发白，我一般跑cfg3.0,30步。请问是提示词有问题吗？哪里可以学习这模型的提示词写法

@mistania base是用来练lora的，我们用sft，然后cfg官方推荐3-6，发灰可以试试提高cfg到5

@wdwhjh 好的，我试一下，我用了sft但感觉AI脸有点严重

@mistania 反面提示词加一些类似 美颜 网红 之类的提示词会好很多，还有就是对面部和皮肤描述多一些