Teatime Dream

TeatimeDream Neo

The Neo series is made from materials that have been fully fine-tuned using images from SDXL and Quen, and while it has some excellent aspects such as the fingers, body expression, and color depth, it had a fatal problem in that it didn't look cute when simply combined. It's not so terrible, but it fell short of my expectations.

This model is a trial merge of TeatimeDream with a model based on materials from 0.1 to 0.4. That said, the main improvement has been in the color, and not much of the essential aspects of the Neo series have been added.

This time, the emphasis was on cuteness, but if the proportion is increased, the product will deteriorate significantly. It is probably possible to further refine the balance between the improved and deteriorated parts, but even with this level of improvement, my original goal of creating a TeatimeDream with improved color has been achieved.

In conclusion, I found that the Neo series can be used to improve SD1.5 models that have issues such as pale coloring and weak body structure. How much improvement can be made without losing the characteristics of the original model depends on the extent to which it is possible.

All samples in this example are output at 600x900 and have HighRes.Fix 1.5 applied. This was added just in case, as the angle of view changes when merging to strengthen the background, causing characters to appear smaller. The LCM output is 7 steps. Using a sampler with a higher number of steps may improve the appearance of fingers, etc. However, using an excessive number of steps will have the opposite effect.

---

HighRes.Test

This is a test to block-merge NAI2 into an existing model to make it compatible with high resolution. Here, NAI2 is block-merged into TeatimeDream.

As shown below (including failed images with consecutive seeds), although it is not perfect, it is difficult to get a large failure even with an output of 1024 x 1536 dots. On the other hand, the composition tends to be monotonous.

(The top one is TeteimeDream, and the bottom one is the High.Res compatible version.)

---

Teatime Dream

This is an SD1.5 anime merge.

I received a feedback, so this time I made it using non-LCM.

Please use whatever sampler you like. The sample uses DPM++. The number of steps should be set at 20, but increase it if the picture is not stable.

I felt that CFG scale 7 was too hard, so I used 6 in the sample.

SD1.5のアニメ系マージです。

ご意見をいただいたので今回は非LCMで作りました。

サンプラーはお好きなものを使ってください。サンプルではDPM++を使っています。ステップ数は20を基本として，絵が安定しない場合は増やしてください。

CFG scaleは7だと硬すぎる感じでしたので，サンプルでは6を使っています。

The concept this time is "don't experiment too much." I've already done a lot of experiments with SD1.5, so I decided to keep it simple. We have narrowed down the number of models used.

As usual, I don't use very famous models. We aim to stably output pictures of sufficient quality with simple prompts, without quality prompts, negative prompts, or embedding.

今回のコンセプトは「あまり実験しない」というものです。SD1.5ではすでに山ほど実験してきましたので，無難にまとめる方針で作りました。使用モデル数も少なく絞ってまとめました。

いつものことですが，すごく有名なモデルは使いません。クオリティプロンプト，ネガティブプロンプト，Embeddingなしで，簡単なプロンプトで安定して十分なクオリティの絵の出力を目指します。

The current issues are as follows

-Eyes are unstable

-lighter color

現状の課題は以下の通り

-目が不安定

-色が薄め

I tried 20Steps for the first time in a while, and it really calmed my hands and fingers.

I see. That kind of thing wouldn't converge with 8Steps. LCM tries to summarize quickly and roughly.

I'm also dabbling with Flux, so I feel like it takes about 10 seconds per photo to be fast.

久しぶりに20Stepsとかやってみましたが，手指などが非常に落ち着きますね。

なるほど。ああいうのは8Stepsだと収束しないものだったわけですね。LCMは荒く素早くまとめようとしますから。

Fluxにも手を出しているので，1枚10秒くらいなら速めに感じられます。