Anon’s Dream Scene is a 2D anime-style scenery LoRA model. It is designed to enhance the visual aesthetics of anime scene while reducing their realistic look.
Anon’s Dream Scene 是一个 2D 动漫风格的场景 LoRA 模型，用于增强 anima 模型的动漫场景视觉美感，并降低其写实倾向。
Trigger word: anon_scene
触发词：anon_scene
You can find the full tag frequencies on my blog. This LoRA is designed to be triggered easily with common scenery-related tags. For examples:
outdoors:
anon_scene, scenery, outdoors, natureindoors:
anon_scene, indoor
你可以在我的博客中查看完整的标签频率统计。该 LoRA 设计为可通过常见的场景类标签轻松触发，例如：
室外：anon_scene, scenery, outdoors, nature（场景、室外、自然）
室内：anon_scene, indoor（室内）
The style is intentionally a little overfitted. You can lower the LoRA weight (<0.7) to get a better balance when using it with other LoRAs.
该风格有意进行了轻度过拟合处理。在与其他 LoRA 叠加使用时，可以将权重降低到 0.7 以下以获得更好的平衡效果。
Style Description 风格描述
Indoors: Bright lighting, clear shapes, and sharp line details.
室内：明亮的光照、清晰的形体结构、较锐利的线条细节。
Outdoors: Bright anime-style colors.
室外：明亮的日系动漫风格色彩。
The model was trained on natural photos, animation screenshots, scenery illustrations, and 3D models with cartoon-style rendering.
该模型的训练数据包含自然照片、动画截图、场景插画，以及卡通渲染风格的 3D 模型。
Description
This version is strongly anime style oriented.
FAQ
Comments (2)
First of all, thank you very much for the work put into this LoRA. Looking at the preview images, I can't help but notice that this LoRA seems to significantly hinder prompt adherence, particularly regarding lighting and atmospheric tags. While the visual output is excellent, the final result seems somewhat hit-or-miss when it comes to following the prompt. I also imagine it alters the artistic style quite a bit, doesn't it?
Thanks a lot for your comment.
About the atmosphere related description, that is indeed the case, and it mainly comes from the tagging and captioning strategy. In this LoRA, object-level elements were annotated, while atmospheric descriptions were intentionally removed. As a result, atmospheric elements tend to be absorbed into the LoRA’s stylistic behavior and are effectively tied to the concept of “anon_scene”(theoretically). You can also check how the dataset was labeled.
Regarding style, as mentioned in the model description, this is an anime-style scenery LoRA model, so it does reduce photorealistic tendencies. That said, LoRA alone cannot fully perform style transfer on a given image. This is my fifth iteration of this model (one of roughly 120 LoRAs I’ve trained). Earlier versions used a large amount of 3D renders and anime backgrounds derived from real photos. Those earlier versions were more faithful to prompts, but the style was less stable : sometimes producing realistic-looking images, and other times achieving very strong anime aesthetics (almost identical to anime base seeds in some cases).
Although many people may prefer a more realistic scene style, that is not the direction of anon_scene.
In comparison, this version is more strictly oriented toward a 2D anime-style scene. I’m still not entirely sure if this is the “correct” direction, but it currently matches my intended design for the LoRA. That said, the balance in captioning still needs improvement. Some less common terms were deliberately cleaned up for consistency, but in doing so, certain distinctions (for example, different types of skies) were simplified too much. Overall, the tagging approach makes this LoRA behave more like a “gacha-style combination of objects” rather than a precise natural-language description of spatial composition.
If the base model has strong enough semantic understanding, lowering the LoRA weight can help restore balance, as also mentioned in the model notes.
Personally, I don’t particularly recommend using anima as a base model for scenes. Its adherence to object-level prompts in scenes is not very stable, and I’ve seen many cases of issues such as missing elements (for example, “rabbit on the moon” without the moon, or a girl with a parasol generated without the parasol). Although the example images mostly use anima for benchmarking purposes, in my testing, switching to other base models consistently improves both aesthetic quality and prompt adherence.
This is just my personal experience.