Comfy-Org

/

MOSS-Image

Like 349

Follow

Comfy Org 15k

lucky for you there are no people with .... scary creeepy music playing... HANDS in it.
This may be worse at people anatomy than flux2/3 klein? I guess it does have other strengths.

lucky for you there are no people with .... scary creeepy music playing... HANDS in it.
This may be worse at people anatomy than flux2/3 klein? I guess it does have other strengths.

Nope, hands seems fine even with a single sentence prompt....... NICE.

1. arguably no. the hand on the left looks already a bit weird even in this simple image.(small finger part too wide) but may just be me.,
   but anyway, even sd1.5 (ca.2022) could get a hand in plain sight right.

The issue is if a hand is involved in any interaction, holding something, touching something. The worst anatomy model in recent history was Ernie image, which actually struggled with a person just standing there. to the level of not getting two legs right. Other models like krea2 or the underrated boogu image rarely make any anatomy mistake, whatever you throw at them (example each linked further down) . While Flux2 and 3 do. And moss is deffo closer to those.
Also it stretches or squeezes people to fit them in toe aspect ration instead of keeping normal proportions (needs a reminder in the prompt).

let's revive an old classic from this flux.2 dev thread, a MASSIVE model that is really bad at anatomy https://huggingface.co/black-forest-labs/FLUX.2-dev/discussions/12 . check there to see it's mess,
This is moss (the EXACT comfy moss t2i template, just my own prompt enhancer (but same system prompt).:

first basic prompt: two young business men in sloppy dress shirt and tie, shake hands on a deal, nearby others are arguing with each other.

enhanced prompt, adds more detail but it does not fix hands. (look at those in the back) This is VERY much like the flux.2 messup linked above:

my silly prompt to test complex scenes. Here is what krea2+kroma1 does, next to it boogu image.
krea2+kroma: https://huggingface.co/Andyx1976/Kroma-v0.1_Lora/resolve/main/images/kroma-_04477.webp
boogu image 0.1 base+turbo lora (works better than the turbo model) : https://cdn-uploads.huggingface.co/production/uploads/67786db89328893864293f2c/9Pr9G5nPekp91Vr57p2yp.webp

this is moss (the text placement is good... The fists are a mess, the arms of the left and right guy disappear..:

i removed the cat from the prompt to make it less complex (i only added it because boogu just wouldn't break. This is a bit better but the fists(?) are still a mess.

in running these examples just now i noticed, it puts a lot of hands in pockets (unprompted). i wonder.... Also it just ignores part of the prompt. lets assume its the prompt formatting. BUT a messed up hand is not caused or prevented by a prompt , that is a training problem. I did not bad-cherry pick, i took the first result where not everybody had his hands in his pockets. It really is that bad (official comfyui moss t2i workflow).

same but with 40 steps:

This is BAD! And it screams AI even with the quite good image quality. I thought we were past "count the fingers to detect ai" time.

1. arguably no. the hand on the left looks already a bit weird even in this simple image.(small finger part too wide) but may just be me.,
   but anyway, even sd1.5 (ca.2022) could get a hand in plain sight right.

The issue is if a hand is involved in any interaction, holding something, touching something. The worst anatomy model in recent history was Ernie image, which actually struggled with a person just standing there. to the level of not getting two legs right. Other models like krea2 or the underrated boogu image rarely make any anatomy mistake, whatever you throw at them (example each linked further down) . While Flux2 and 3 do. And moss is deffo closer to those.
Also it stretches or squeezes people to fit them in toe aspect ration instead of keeping normal proportions (needs a reminder in the prompt).

let's revive an old classic from this flux.2 dev thread, a MASSIVE model that is really bad at anatomy https://huggingface.co/black-forest-labs/FLUX.2-dev/discussions/12 . check there to see it's mess,
This is moss (the EXACT comfy moss t2i template, just my own prompt enhancer (but same system prompt).:

first basic prompt: two young business men in sloppy dress shirt and tie, shake hands on a deal, nearby others are arguing with each other.

enhanced prompt, adds more detail but it does not fix hands. (look at those in the back) This is VERY much like the flux.2 messup linked above:

--
my silly prompt to test complex scenes. Here is what krea2+kroma1 does, next to it boogu image.
boogu image 0.1 base+turbo lora (works better than the turbo model) : https://cdn-uploads.huggingface.co/production/uploads/67786db89328893864293f2c/9Pr9G5nPekp91Vr57p2yp.webp

this is moss (the text placement is good... The fists are a mess, the arms of the left and right guy disappear..:

i removed the cat from the prompt to make it less complex (i only added it because boogu just wouldn't break. This is a bit better but the fists(?) are still a mess.

in running these examples just now i noticed, it puts a lot of hands in pockets (unprompted). i wonder.... Also it just ignores part of the prompt. lets assume its the prompt formatting. BUT a messed up hand is not caused or prevented by a prompt , that is a training problem. I did not bad-cherry pick, i took the first result where not everybody had his hands in his pockets. It really is that bad (official comfyui moss t2i workflow).

same but with 40 steps:

This is BAD! And it screams AI even with the quite good image quality. I thought we were past "count the fingers to detect ai" time.

Expand: the boogu post has some other examples like qwen image and this apocalyptically (!) bad flux2. klein 9b one. https://huggingface.co/Comfy-Org/Boogu-Image/discussions/9#6a6700d292352a336441fb4e

No problem with the model—it’s incredibly good (that’s how Flux3 should have been). It’s a prompt issue; the English you have there is crazy. If it helps, here’s an image—there are others with no issues with the hands or the body. I admit it wasn’t easy; prompts in English are always crazy, but with a little tweaking and good writing, the problem goes away. Some models are more prone to having poor grammar ruin complex parts like fingers, hands, and eyes. It’s not a problem with the models—it’s just that AIs speak better than we do, and we have to match that, also, they are becoming more like tools for editing and creating images than a model who can paint with minimal guidance, needs direction, and needs to stay focused—not a painter but a tool for generative "paint". This one was done without a prompt rewriter(I have more and no issues); you can see the workflow in the image—simple and straightforward, same seed. Pay attention to the resolution: nice numbers rather than perfect megapixels help the model use the space better. (Placing the cat on the biceps requires a different image composition and more instructions to define spaces; you can use bbox if you read the model’s instructions and compose exactly what you have in mind—just typing “1girl” isn’t enough with these new models anymore ;D). Best regards.

Prompt:

A photograph, 3:4 vertical composition, realistic, detailed and casual styled.

The image composition:
At a workshop. Its background is large and messy, however it remains well equipped. On the back wall, occupying the upper side of the image there is a large dirty poster which has a large, red colour lettering reads "DON'T" while the O on DON'T is replaced by a cog symbol; and below it a slightly smaller and blue lettering reads "Overengineer".

Below the poster, occupying the remaining three-quarts of the image, and in the foreground, there are three men; the man on the left, the man on the right and the young man in the centre.

Main action:
The men on the left and right are each holding and lifting the man in the centre by one arm; while both lifting him off the floor, his arms are stretched outwards and upwards, leaving his hands and fingers loosely hanging down from above behind them in a realistic and natural pose and also covering partially with his head the poster of the background. The man in the centre seems embarrassed as he gazes at the man on the left, who looks sternly back at him.

Subjects:
The man on the left (25, tall, slender, wiry, bald shaven) wears a tight elegant blue dress shirt with black/gold striped tie and a small golden "boss" print on the shirt pocket.

The man on the right (tall, slim, age 30, hairy body, short disheveled black hair) wears a dirty white ribbed tanktop under a dirty blue workers dungaree with a small white "NO BOSS" print on the front; on his biceps, a tiny baby tabby kitten balances.

The man in the centre is young man age 18, average height, wearing blue jeans and a white T-shirt, with a large abstract cogwheel symbol on its front, accompanied by the words "WE engineer" above and "Everything" below; The T-shirt is too short. The lift causes to him to expose his midriff, his low Levi's jeans' beltline, and the orange band of his underpants with the partly visible blue print "Kelvin Cline".

Art direction:
To position the three subjects with their heads at nearly the same height while the two men lift the young man in the center, it is necessary to maintain the frame: the man on the left and the man on the right are shown mid-body from their waist, the man in the centre is shown from knees. Maintain realistic poses without physical distortions; use the space in the frame naturally; If something is obscured in the background, that's preferable to losing the action and the poses. Use the physics of materials and textures to create an accurate and well-defined image, Pay close attention to the positions and shapes of the fingers and arms without distortion and use realistic, detailed poses and grips; Pay even close attention to the fingertips anatomy. The focus is on the young man being lifted, the gazes, the outstretched arms, each man’s biceps pushing up, and the cat looking curiously, which adds a touch of humor as it perches calmly on one of the biceps.

no. a hand is a hand. If a model needs a enhanced prompt to make a hand it is NOT "good" (at people).

This is worse than Flux3, it is probably worse than flux2 9b in complex scenes. In simple scenes flux2 9b will get it right while moss will mess up every hand in a vaguely challenging position. A hand of a person in the background, let alone the center has to look like a hand in 2026.
I mean look at krea2 and boogu image example. I actually tried, your micromanagy prompt suffocates both models into a worse result (still ok hands).

You don't need to write a fricking novel to put three people in a image. With NOT worse results. OR you need a chapter on what a hand looks like.
The real issue is: Why use a model that needs a book when others can just do it.

The issue is not the composition. It understands what is wanted. In your prompt and in mine. That is perfectly ok in all examples. It just can't do hands like its 2023.

but why not. your prompt (load it in comfy): Even in good attempts there is something wrong with the hands.

this is flux1 dev (or srpo) with my bad prompt. it's a redux workflow, that's why i still have it, but every input but the prompt was bypassed. (i wanted to show of sdxl but that was a bit of a stretch)

Yes, I get your point. For krea2 and boogu, you need a different kind of prompt, of course; that one overwhelms them, and they can’t handle that many tokens. This model accepts 4k prompt tokens, but it’s very sensitive to grammar, like others in its class ,and errors accumulate during inference. My point is that this model isn’t bad at anatomy; it’s just designed for more than just “fill in the blanks.” What gives it precision with bounding boxes and hierarchical prompts, just like other new models that have been coming out, also it takes away its freedom to “paint” on its own and fill in the blanks the way people have gotten used to. You said it yourself: if that’s what you want, use a different model; but if you want a precise, Photoshop-style editor, use this one or others with the .json system and flawless prompts. They’re two different things, some people paint; this is a tool. But don’t deny that a mediocre prompt doesn’t really put anything to the test, it just highlights that a model will almost certainly fail, and that, with all due respect, as constructive criticism, is useless in comments; even more so since it’s free and based on the work others have done to make it possible. It’s a matter of being objective; I presented my opposing view so as not to discourage the reader from understanding the conditions required to achieve it. If that offended you, I apologize. Best regards.

Less micromanagy prompt suffocation, just a couple of tweaks (notice I didn't use negative at all):

Empujando al modelo a que se líe con movimiento, con más práctica, las manos borrosas se veran correctas: