Ah, I see! If you actually want to control the darkness and prevent the background from getting washed out or ultra-bright, you need to guide the model using natural language contrast and specific lighting cues in the positive prompt, rather than just relying on negative tags.

Newer models with vision encoders (eg. Qwen) respond amazingly well to descriptions of atmosphere, light, and shadow. By explicitly describing the interaction between deep darkness and dramatic light within a natural sentence, you can achieve a beautiful night scene without the AI overcompensating and making the background too bright.

Here is an example of how to guide the model for a dark, high-contrast, atmospheric scene:

Positive (Tags + Natural Language):

Tags: 1girl, solo, flower, pink_eyes, medium_breasts, from_side, pink_hair, twintails, long_hair, ahoge, dress, socks, brown_dress, bow, black_socks, sleeveless_dress, hair_ornament, hair_flower, hair_bow, night, night_sky, sky, sitting, looking_up, outdoors, white_flower, cleavage, bare_shoulders, breasts, flower_bracelet, building, lamppost, kneehighs, spaghetti_strap, black_bow, girl, tense, fierce, cold_palette, stark_contrast, shadows

Natural Language: A solitary girl sits quietly in the night outdoors, her vibrant appearance creating a striking contrast against the cold, harsh environment. She gazes up at the night sky, her expression filled with intense tension and fierce emotion. Her vibrant dress and elaborate hair ornaments vividly oppose the stark shadows that envelop her. A cold palette of light dramatically highlights her skin and attire, emphasizing a sense of cold beauty. The juxtaposition of deep darkness and brilliant light casts dramatic shadows, intensifying her delicate presence. The scene is permeated by an atmosphere of stark contrast, capturing a moment where intense emotion and cold reality collide.

Negative (Keep it minimal):

worst_quality, low_quality, bad_anatomy, extra_limbs

Key Takeaway:

By using phrases like "stark shadows that envelop her", "juxtaposition of deep darkness and brilliant light", and "cold palette of light", you give the model the exact context it needs to render deep blacks and sharp highlights naturally, instead of fighting against a bloated negative prompt. Try mixing descriptive prose like this into your prompts!

@OimoJet that's quite a huge wall of text. I already knew all of this thank you. but still that doesn't help much. the base model like to put windows inside room, or open space. even with darker tone, the model struggle with darker/somber ambiance. i first thought i was doing something wrong with the model because i never had such problem before, after digging some info, it found out it was the model itself. in my negative i use " overly bright palette:1, washed out colors, overexposed, flat lighting, bright background:1.5", it help a little bit with some models, but Hakushi Mix Anima is incredibly hard to not get overly bright colors. it's not impossible but hard, but even the model (anima base) will try to push it toward something grandiose with a scenery. if you look the picture posted down bellow from people with this model you can clearly see what i'm talking about. one "trick" i found that work it's to find an artist that do not use harsh colors, this really help and change a lot the picture.

regardless, thanks for your reply.

@tonikakushiroi784 hi, i rarely use Euler A, i use mostly Er_sde - Beta, i tested many sampler, for CFG i use 4 to 6, depending what i'm doing. I can get good result with other models, but with this particular model, it's very hard. I also use comfy.

@tonikakushiroi784  thank to you too, taking the time to answer is really appreciated. >⩊<

попробуй для затенения: https://civitai.red/models/2526312/darkness-slider