v4.3 [final]
I recommend to choose this one or v2. If you don't care about speed and want great anime results use this model + LCM sampler + 20 steps. It's simply better than any model I know in this aspect.
Recommended settings:
6+ steps (8 + 4 steps of upscaler is optimal for my use cases, feel free to experiment).
A few options of samplers:
CFG 1.8 | Euler A
CFG 2.0 | LCM - better for anime and other abstract things.
CFG 1.8 | DPM++ 2M SDE with BETA scheduler (normal doesn't work)
Use style tags such as "anime style, flat coloring" or artist/title name. It will help model to know what you want from it.
The fastest generation: use CFG 1.0 and set lora CFG_slider on negative -0.4 value. Lora listed in suggested resources. (CFG 1.0 ignores negative prompt. Don't use it if you need negative)
The best coherency: use CFG 2.8 and CFG_slider lora on +0.75 weight (for Euler A).
v3.2B
Recommended settings:
CFG 1.8
LCM / [kl optimal / sgm uniform]
5-10 steps
v3.2A
Recommended settings:
CFG 1.8 (works not bad on cfg 1 as well)
Euler A / normal
5-10 steps
v1.99 (v-prediction)
Use title or artist name to get a specific style.
Eauler A - Beta scheduler (or normal)
8-16 steps
CFG: 1.0 recommended for 2x speed. Use more if you want to use negative prompt.
v2
Sampler: Euler A / Euler A cfg++
CFG: (depends on sampler)
Alternative sampler: LCM
Description
Alternative which potentially may be more creative and have better compability with LCM sampler.
FAQ
Comments (4)
The author mentions the recommended environment, so this may not be necessary, but I looked into it and will write about it here.
The test environment is v3.2B.
Number of steps: The model says 8 steps are OK, but it seems better to have more steps. In other words, there is no need to stick to 8.
Scheduler: In the test environment, you can use anything other than KL Optimal. I think beta/turbo/SGM Uniform are good, but if it's too much trouble, auto is fine.
CFG:1 gives a dark impression, so I'd like it to be around 1.5. The author recommends 1.8, but personally, I recommend 2. Above 2.5, the colors become distorted, so within the range I tested, it needs to be 2.1 or below.
Sampler: Since there is behavior where “the number of steps is traded off for distorted images or colors” as seen in acceleration LORA, I recommend the recommended LCM or Euler a. Euler a is recommended for V3.2A, but I think V3.2B is also acceptable.
推奨環境について作者が述べているので不要かもしれないが調べたので書いておく。
テスト環境はv3.2B
ステップ数：8ステップでOKというモデルだがステップは多いほうが良さそう。逆に言えば8に固執する必要はない。
スケジューラー：テスト環境だと、KL Optimal以外はお好きにどうぞだ。beta/turbo/SGM Uniformあたりが良いと思うが面倒ならautoでも良い。
CFG:1は暗い印象なので1.5くらいはほしい。作者の推奨1.8だが個人的には2がおすすめ。2.5を超えると色がおかしくなるので試した範囲だと2.1以下にする必要がある。
サンプラー：高速化LORAなどに見られる「ステップ数と引き換えに絵や色がおかしくなる」挙動があるので推奨のLCMやEuler aがオススメ。Euler aはV3.2Aの推奨だがv3.2Bでも良いと思う。
Honestly, I'm not satisfied with v3.2 of this model. It looks like big portion of blocks from Illustrious 2.0 ruined it. It have a good character recognition and precision, can easily switch styles, but it's not as good as v2 for short prompts, where I want a more randomized and creative result. I will make something better...
I would recommend you to try v2 if you are not tried it yet.
Shio_N I didn't think V3.2B was a bad model in terms of performance or behavior.
I'll try v2 too.
V2
Although we have prepared samples (comparison list), we recommend searching for all patterns based on user preferences.
There are probably not many models that vary this much with the sampler.
In that case, please check the scheduler as well. It will vary.
The results differ between 8 and 12 steps, so please check this as well. Probably 8 is the default.