lodestones

/

Kroma

like 287

Alright after much testing, I have found some strong points and some other things.

1. It is much better at compositional and pose understanding. (A little bad in terms of anatomy and features detail)
2. Much better as general model with capability to follow many styles.
3. Can do much better with small prompts, compared to large prompts.
4. Knows a lot of NSFW concepts by their common terms, no need to explain position for each body part.
5. Lacks in visual clarity (By visual clarity I mean textures and edges e.g Rough cartoon outlines, skin texture, ragged eyes and hair, blurry edges). Using style or character LoRAs trained on Krea RAW base model, cripple the visual clarity further. Maybe it needs LoRAs trained on Kroma Base. (Can't train using my hardware so couldn't test this)
6. Lacks in emotions, although Krea itself suffers from it but in my opinion it loses some further compared to base Krea too.

Overall I feel like its very capable as an early v0.2 model which hopefully is scheduled to get better! Cheers to lodestones, great work!

I downloaded the model but I couldn't find what strong points it has compared to Krea2+Some NSFW LoRAs or other uncensored merges available on Civitai. Its a genuine question so please don't get agitated, I personally made a merge based on Krea2 RAW with Turbo LoRA and some other NSFW LoRAs and my merge gave me better results. Maybe its better but my prompts are not?

The merge I am using:
Krea2_TextFusion_Refusal_Reduction.safetensors: 1.0
https://civitai.red/models/2775340/krea2-textfusion-refusal-reduction-lora

fedor_bypass.safetensors: 1.0
https://civitai.red/models/2746817/krea2-filter-bypass-fedor

krea2_turbo_distill_r256.safetensors: 0.6
https://huggingface.co/TheDivergentAI/krea2-turbo-distill-lora/blob/main/krea2_turbo_distill_r256.safetensors

snofs_krea_v1_3.safetensors: 0.8
https://civitai.red/models/1972981/snofs-sex-nudes-other-fun-stuff?modelVersionId=3104629

KreaAmateur_V2.safetensors: 0.2
https://civitai.red/models/689192/aesthetic-amateur-photo?modelVersionId=3180950

skindetails_krea2_loraholic.safetensors: 1.0
https://civitai.red/models/2682644/skin-detail-slider-krea2-zit

Breast_Nipple.safetensors: 0.2
https://civitai.red/models/1259365/ultrareal-breast-and-nipple-detailer?modelVersionId=3146215

Krea2-brs-sqz-V1_epoch_10.safetensors: 0.3
https://civitai.red/models/2761661/grabbing-breasts?modelVersionId=3161094

Settings: 12 Steps - ER SDE - Beta - CFG 1

This happens because Kroma is supposed to be the cleanest unlocked version of Krea 2. For me, Kroma is the best finetuned version of Krea 2. It's practicality is incredible, no need a bunch of LoRas and performs very well as the base model, and you also get new concepts for some stuff. That's why I wanted a Chroma version of this model as well.

@Shiza001 Do you mainly do cartoon/anime stuff, or is Kroma actually great for natural realism as well?

@Shiza001 Do you mainly do cartoon/anime stuff, or is Kroma actually great for natural realism as well?

I do both. Kroma works great for realism as well.

The lack in visual clarity is probably due to the training being on early stage and trained for low resolution only.

@Shiza001 What sampler/scheduler are you using?

After some testing I have to agree: Kroma works great for realism.

Update: The below was written on kroma 0.1 lora+krea turbo. In that aspect of following complicated prompts, Kroma0.2 seems to be a massive step BACK. Only tried the turbo model. Even the extracted lora, someone posted combined with Krea2 messes stuff up badly.
Putting the "old" Kroma0.1 rl mild lora on the same krea model gives a vastly superior (not messed up) image in terms of composition. The visual image quality (textures and stuff) maybe better but that doesn't help much.
So if you're unhappy with the full model Kroma0.2, Give Krea2 turbo with Kroma 0.1 a go! But Kroma0.1 or Kroma0.2 rl mild. The bigegr one Kroma0.1 RL was for me basically fully censored.

old post, based on KRoma0.1+Krea2 turbo
it gets really complicated prompts with many people,. Not just plain nsfw body details and the most common use cases, that many loras cover, but also almost every composition, kinky situations, with multiple people, different interactions by different groups in the same image, others in the background, and with placed text and so on. All the details you prompt are there, and the anatomy in question STILL works flawlessly. As does the normal anatomy.
And it just does it. Almost physical situation you throw at it. Very rarely messed it up things like arms and legs. It kinda replaced Boogu image for me in that, which is also excellent at these complex scenes, but it is not uncensored.

I don't know how much of this is Krea2 (when uncensored). Text placement for example certainly is helped by the excellent (if free'd) Krea2 underbelly. But instead of having multiple loras for multiple ... cases, Kroma does it all out of the box and very nearly perfectly, whatever you throw at it. A simple anatomy extension lora can NOT do that.

Also unlike the VAST majority of all-in-one nfsw models Kroma handles non male-female situations flawlessly. Without you having to hammer the point home in your prompt (lustify!) , without the lora/model crowbaring unprompted features in. That's quite annoying for example with Dasiwa video lmodels. Where it happens subtly but it does. Kroma (and Chroma1HD btw) makes what you prompt.

Let's interrupt the tiptoeing for a moment: For gay stuff this is the only general all-in-one Image model i know that JUST does what you prompt. If you prompt guys, you get guys. I strongly assume it's the same for ladies. And it remains as capable and versatile as it is with the usual combo.
Because as someone up there says it just knows the words, and terms, in whatever combo.

you don't have to trick a vaguely de censored model into doing it. You don't have to remember 10 trigger words. Just a natural sentence prompt, or even a more cryptic, listing all the details,, all the people and groups of people and you get it. I think people got really obsessed with the often unneeded short story prompts, you may not do Kroma a favor with a flowery Z-image prompt, but that is another topic.

Yes, the visual clarity, skin details and stuff need some work, although samplers can help with that. And what use is nice visuals if 2 out of 3 images mess up even hands and arms. Then it really doesn't help if a lora add the other stuff.