The P series was intended to be more photorealistic than the R (whether or not this was achieved is another matter).

P5 aims to be more realistic than R2 while incorporating fantasy elements, but when you add "fantasy", it has the drawback of becoming illustration-like. General prompts produce photo-like images, and can also produce NSFW images to a certain extent.

P6 is an even more photo-like version of P5, and can produce photo-like images to a certain extent even in fantasy. However, the prompt tracking and facial expressions are far inferior to P5, and although it can produce beautiful images, it may lack interest.

Naturally, I also created a block merge model that would produce P5 images in the P6 style, and initially planned to release only that, but after looking back, it was a rather half-baked and uninteresting model, so I decided to release both of them.

I'm not used to Flux block merging, so it won't work like SD1.5. I'll study it a bit more.

Finally, my first Flux.1 model.

To be honest, I think it's wise to use only standard models in Flux and handle the rest with LoRA, but I decided to make a merge model including investigating the internal structure.

However, there are too few materials. In addition, the model creation environment is not at all ready. After a lot of struggle, I gave up on Forge and Supermerger and somehow managed to put together a spaghetti workflow with Comfyui. There are still limitations to what can be done compared to the SD1.5 tools, but I would like to try various things within the current range. By the way, for the time being, I plan to only handle Flux.1 Schnell.

ようやく私の最初のFlux.1モデルです。

正直な話、Fluxでは標準モデルだけを使い、あとはLoRAで対応するのが賢いやり方だと思っているのですが、内部構造の調査などを含めてマージモデルを作ってみることにしました。

しかし、素材が少なすぎます。加えてモデル作成環境が全然整っていません。私はいろいろ苦闘した挙句にForgeとSupermergerを諦めて、Comfyuiでなんとかスパゲッティなワークフローを組み上げてみました。SD1.5のツール群と比べるとまだまだできることに限りはありますが、現状の範囲でいろいろ試してみたいと思います。ちなみに、当面はFlux.1 Schnellのみを扱っていく予定です。

FluxTest R0002

As usual, I tried to create a fantasy-oriented model. Initially, I had planned to make it a little more realistic, but due to the materials and other circumstances, the output was more like an illustration.

いつもの路線といえばそうなのですが，ファンタジーを指向したモデルを作成してみました。当初の予定ではもう少しリアル寄りにする予定だったのですが，素材の都合や諸事情により，まさにイラストレーションという感じの出力が多いものになりました。

Steps are expected to be around 6.

For some reason, it is not possible to incorporate a custom node for adding metadata to images, so most of the samples are produced using Forge. Many of the image scripts are the same as those from the SD1.5 era, and the prompts are not optimized for Flux.

Stepsは6程度を想定しています。

なぜか画像にメタデータを入れるカスタムノードを組み込めないので、ほとんどのサンプルはForgeで出しています。多くの絵のスクリプトはSD1.5時代と同じもので、プロンプトはFlux用に最適化されているわけではありません。

The image style is not stable. Well, I used a lot of prompts that did not specify a style, so it would have been strange to ignore the specifications and fix the image as in the previous models, but the image style is different for each seed? This was not possible in SD1.5, so it is quite refreshing. I wonder if the image is generated by a method other than denoising.

絵柄は安定していません。まあスタイルを指定していないプロンプトを多用しているので，これまでのモデルの，指定を無視して絵柄を固定していたほうがおかしいのだと言えばそれまでなのですが，シードごとに絵柄が違う？　というのはSD1.5ではありえなかったので，なかなか新鮮です。デノイジングではない方法で画像が生成されているのでしょうか。

The model is created by creating one workflow for each model. This is almost my first time using Comfyui, so I may be doing something inefficient. This time, the model is created based on flux1-schnell.safetensors, so the base model seems to have a strong influence on the lack of realism. I can't express NSFW very much. My plan is to change the base part of this workflow to make another model more realistic and fantasy-like.

モデル作成は1つのモデルにつき1つのワークフローを作ることで行っています。Comfyuiを使うのはほぼ初めてなので，効率が悪いことをしているかもしれません。今回は，flux1-schnell.safetensorsをベースにモデルを作成していますので，リアルになりきらないのはベースモデルの影響が強そうです。NSFWな表現もあまりできません。目論見では，今回のワークフローのベース部を変えれば，別のモデルをもっとリアルファンタジー系にできると予想しています。

Flux often has a rubbery or airbrushed skin texture, so I tried to avoid that as much as possible by leaving a little noise, but this doesn't work well with quantization, so I had to tone it down, and personally I think the result is a little disappointing.

I've also tried to tone down the faces, which tend to be incredibly beautiful, and aim for a more natural look for both the face and the skin.

Fluxではゴム調というか，エアブラシで描いたような肌の質感が多く見受けられるので，それはなるべく避けるべく，少しノイズを残すようにしているのですが，これが量子化と相性が悪くて控えめにせざるをえず，個人的にはいまひとつな仕上がりになりました。

もの凄い美人顔になりがちなのも少し緩和して，顔も肌もナチュラルな表現を目指しています。

The drawbacks are that the background objects are noisy and the monsters are too detailed. I know that this is a side effect of putting in a lot of different things, but it's difficult to adjust. SD1.5 had a lot of adjustment options, so it was possible to deal with some of the issues, but Flux doesn't allow for that.

背景のオブジェクトがうるさい，モンスターのディテールが細かすぎるみたいなところは難点です。このあたりはいろいろ詰め込んだ副作用だとわかっているのですが，調整は難しいです。SD1.5では調整要素がたくさんあったので少しは対応できたものが，Fluxではできないからです。

By the way, FluxTestR_00002 means that there is also 0001, but it's basically the same line, and it was canceled because it had too much noise, and adjustments were made to get 0002. For the time being, this is an experimental work to get used to Flux (although there are rumors that all of them are experimental), so please note that there may be skips in the numbers due to management reasons.

ちなみに，FluxTestR_00002ということは0001もあるわけですが，だいたい同じ路線のもので，ノイズを乗せすぎたのでキャンセルとし，そこから調整を加えたのが0002となります。当分はFluxに慣れる実験作ですので（全部実験作だとという話もありますが），管理の都合上，今後も番号が飛ぶことがあるかもしれませんのでご了承ください。