Recommendations

Do you not have a problem with the way it tends to lean towards messy sketches and paint styles?

@nvo76 yes, I do have the same problem. I believe is because we need to be very precise with our prompt. artist tags are very important too. This will be better once we get fine tunes /mixes of the model, I think...

@Rehvka

1. WAI loras arent compatible with this model....well uh it's more like it's isnt compatible at all with any other model beside WAI, training loras on a merge model instead of the base model is bad practice (unless you need the lora only for that specific model). Defaulting to sketchy style is a trait of all base models (Noob, Illust, Neta, Anima, Newbie,...) on short prompts, the only reason most dont notice is either due to using artist tags or using a merge model.

2. This model doesn't perform too good on really short prompt so yes you do need to prompt properly with danbooru tags and at decent length (like 8+ tags and above, with background and character traits n all that.)

3. The model is compatible with all NoobAI EPS and (normal) Chenkin Noob loras since they share the same lineage, Base Illustrious loras will also work although weaker in effect.

Bonus: I'm the guy that genned alot of the showcase images and honestly the anatomy is my biggest issue with the model lol, glad you don't think so though. The model actual standout potential is the absolutely superior color depth/dark lighting capabilities, better prompt following and quicker convergence for training.

@RicemanT Yeah, I think that's what a lot of people are missing. Rectified Flow is in its early stages, and it's hard to explain to people why they should use it. The sad truth is that Flow helps the trainer more than the user. It converges faster and is more accurate to the data (though that also means it more faithfully reveals dataset artifacts). But I don't think people really care about that; they just want better images.

@RicemanT 
1- most of my loras are failures so I normally use whatever people post online and some of them are trained in WAI unfortunately. That is something that Im gonna miss moving forward, maybe a lot of community resources might be unnusable eventually.
2- and yeah, that why I say stupid/casual users like me need to start learning how to prompt. No more: "1girl, large breasts, long hair" prompts :^) . The results are worth.
Well, NoobAI anatomy was already pretty decent and in the short time I tested it (since it just came out last night on a work day) I found no mayor issues with anatomy and its a 0.3v so Im assuming things will improved until we reach 1.0
you guys are doing an amazing work, looking forward to your proyect!

@Rehvka I think the best way to learn how to prompt these models is to look at gel or dan and look at how they format tags and what tags, its the best way to start if you dont want to use natural language/hybrid, just find a image and copy the tags and gen with those and see what you get! you learn way more about how the model gens that way

@Rehvka the showcases image all have metadata (prompt and negs), so you can just look at how to prompt directly by using the images we genned lol. It's really not that hard, I get so lazy sometime I even just steal the tagging from a danbooru image instead of writing it myself.

@Hysocs "Flow helps the trainer more than the user" - That's just nonsense. Any modern dit model is flow based: Flux, Zit, Zib, Chroma, Wan, Neta, Cosmos Predict etc etc.. Its simply the better noise prediction method.

@deitychaser Listing flow-matching models proves my point: it’s a behind-the-scenes training optimization, not a visible upgrade. That’s why the community still uses older models. Flow isn't bad, but model size matters way more than sampling math in the model your listing a 12B model beats a 2B regardless of math. You can't credit flow just because a model is big and good.

@Hysocs Math that gets noise to image is literally what this is all about buddy. Saying it in a diminutive way is hilarious. Flow matched sdxl beats vpred and eps sdxl. I didn't quote bigger models to say that the noise prediction method is the main factor here, why would you even assume that I would do that lol. Anyways, whatever.

@deitychaser i dont understand what your trying to make a point against? I think we're talking past each other, I never said flow was bad. My point was just that right now, it's a training benefit rather than a direct inference one, especially since rect flow SDXL is still brand new. In the context of a broken LoRA, I was explaining why you'd even use the model yet (because it trains better and thats a good thing for everyone due to flow directly). As a model trainer, that's my bias, so sorry if it came across as me calling flow inferior.

@Hysocs  The intention for rectified flow was never to make training easier for these small sdxl teams but to improve the performance of the sdxl architecture for the consumer. Sdxl is small and therefore easy to train no matter what math you use for your noise, so the efficiency argument that you applied to larger models has really nothing to do with the decision to go for flow matching. Interpreting it as a obstacle for the user might be your subjective impression but nothing beyond that. Yeah, it benefits training. That doesnt lead to the conclusion that the training benefit was the main reason why they developed it for sdxl. And if you mean by that: It makes it easier to predict an image closer to the "base truth" of the dataset - well this directly means that the consumer ends up with a more precise model, which in the end is what we want, right? You can't separate these thing and play against each other.

@deitychaser i think we've been arguing past each other the whole time — i was never making a claim about flow's technical merits, i agree it's the better method. my original point was purely about user perception. most people in this community don't know or care what noise prediction method a model uses, they just want good outputs. that's a behavioral observation, not a technical one, and nothing you've said really contradicts it.

I was talking specifically about SDXL rect flow, not flow matching as a whole. Your counter-examples don't make sense either: you listed large DiTs, but they are fundamentally different architectures and sizes, so you can't compare them to SDXL. Finally, you're misrepresenting me. I said rect flow helps trainers a broad fact thats true in general as if it didnt help training then trainers wouldnt use it, and you responded as if I claimed flow wasn't a good noise predictor or that it was directly worse than something, which I never said.

this is becoming word slop over a argument that has no good points, you are comparing apples to oranges. so if you want to continue just dm me

what you are saying or what im getting at is "If it makes the model better, then it helps the user the most." and all i was saying is that the general user doesnt understand the math so it helps the trainer make better models in return helping the user

@Hysocs I listed the models because I thought you are not much experienced and thus named models purely on the basis that they are more modern than sdxl to point out that modernizing sdxl in that way makes sense. I dont see any contradiction here.

I'm not comparing apples to oranges, i gave an easily accesible example to draw a conclusion.

I assumed you are this hypothetical "genral user" you are talking about, thus i listed examples for "general user".

@deitychaser I offered DMs because this is getting circular.

You admitted you assumed I was inexperienced — that assumption was wrong, as I train rect flow models and open source the training code (check my profile).

The contradiction still stands: you cited 12B-parameter DiT models to argue about SDXL flow matching. That's apples to oranges.

Flow helps trainers build better models, and users benefit indirectly, the two are intertwined, not exclusive. That's not a criticism of flow, it's an observation about who actually notices the difference.

In direct relation to SDXL rect flow specifically: at the time of my comment, it doesn't beat base vpred or eps on quality, but it does win on knowledge. The trainers themselves have stated multiple times that there are known quality issues and it needs more training.

You're arguing with someone who's on your side. The distinction I was making was general public adoption vs. training, two different conversations.

My original comment was in response to an SDXL poll where 48% of users still voted for Pony SDXL, an eps model. That's what I meant by flow helping trainers more than users right now, I was talking about adoption, not technicalities. If users truly cared, eps models wouldn't have thousands of weekly downloads while SDXL rect flow sits at hundreds, if that.

If you want to see how rect flow actually works in the training code, my DMs are open.

Otherwise, let's leave this thread for people who actually want to use the model — not flood a month-old comment.

TLDR; im saying Flow is good but users dont know that yet due to only like 5 models total existing (meaning no loras) so adoption is low. at this point training is for the sole purpose for the trainer to train easier to make a better model not for adoption.

In addition, the color overflow and contamination of 0.3 seem to be more severe.

@bluvoll I've used it in comfyui but the generated image is just noise and black background

@thegoatedone1 bro is a dumbass