Artium_anima版本v3.0 (ID: 432242)

I really, really like the Artium v1.0 Model (credit to FrenzyX).

But it sometimes kind of drifts off into photorealistic output, which is not what I desire from it, so I decided to make a merge variant that would mostly behave like Artium but be more likely to fall back to anime style than photorealistic.

In the end I decided to merge in some amount of the anima_pencil-XL checkpoint, which itself is a merge of blue_pencil-XL and Animagine XL.

Regarding permissions: I'm very much an amateur at this, so I just copied the more strict set from Artium.

Regarding prompts and generation settings: Up to you really.

I like to keep my prompts simple - see the sample images I uploaded. No need to overcomplicate things anyomore.

I mostly use DPM++ 2M Karras with 20 steps, but that's a matter of personal preference and because these settings will work well with most models.

I did a bit of trial and error before finding a merge recipe that seems to have achieved my intent.

One of the experiments I did was take the weight differences between Artium v1 and SDXL and add 50% of that back into Artium for what I called Artium 150%. This resulted in a more anime-like model that near enough achieved all I wanted on its own without breaking.

I also tried adding the weight differences between animapencil and sdxl into Artium, but the outcome had more more flat tones and less texture than where I was intending to go.

The end model I uploaded here is a merge between those two models and another 10% of the weight differences between Artium and SDXL added in, giving the model effectively 135% of Artium and 25% of Anima_Pencil.

I compared about 250 generated images between my model, Artium v1.0 and a 20% Weight-Sum between Artium and AnimaPencil to make sure I didn't break anything while I was at it. Looks good.

About 50 of those images were specifically to check if hand generation got worse - which I'm happy to report it didn't. There were no statistically different outcomes looking at 90 different hands per model.

描述:

Yeah, I skipped v2.0

There's an internal 2.0, but it didn't achieve what I set out, being pretty hardcore flat colour anime. There's plenty of good models for that already.

Overall making of:

Artium 150%: Artiumv1 + (Artiumv1 - SDXL)*0.5

-> much less likely to default to photo style

Artium_Anima - V0.1: Artium 150% + (Anima_Pencilv1 - SDXL)*0.5

Artium_Anima - V0.2: 0.5*Artium_Anima - V0.1 + 0.5*Artium 150%

Artium_anima - V1.0: Artium_Anima - V0.2 + (Artiumv1 - SDXL)*0.1

Prep1: BluePencil_v401*0.5 + AAM XL*0.5

Artium_anima_v1.2: ArtAn+AAM+BP

Artium_anima_v1.3: Artium_anima_v2: +quietGN_AD20

-> very good version, but didn't seem like a big enough change to publish

Prep2: BluePencil_v5*0.5 + AAM XL*0.5

Prep3: Artium 150%*0.5 + Prep2*0.5

Prep4: Anima_Pencilv2.6 + (quietGN-SDLX)*0.6

Artium_anima_v2: Prep3*0.6 + Prep4*0.4

-> Too much detail lost, flat colours, v1.3 is better

Attempts to improve AAv2:

AAv2.1: AAv2*0.5 + AAv1.3*0.5

AAv2.2: AAv2*0.8 + Av1.0*0.2

AAv2.3: AAv2.1*0.5 + AAv2.2*0.5

-> none are good enough for what I'm trying to achieve

AAv2.4: AAv2.1*0.75 + Av1.0*0.25

AAv2.5: AAv2.1*0.75 + A150*0.25

AAv2.6: AAv1.3*0.75 + AAv2.2*0.25

-> nearly indistinguishable from each other. Pretty good - I ended up using AAv2.6 for a while

AAv3.0: AAv2.6*0.6 + randomaxxArt*0.4

The one I published here. I quite like the Randommaxx Art Merge.

训练词语: