After a month of heavily testing single-character generation and migrating my previous workflow to be Anima-exclusive, I wanted to share my findings. Anima-base v1.0 handles concepts that used to require LoRAs or ControlNet through prompts alone.
Please keep in mind that I am still actively exploring the model, so my findings might not be absolute and my methods will likely evolve! Below is a breakdown of what I tried, what worked, and exactly how I structure my prompts to force the model into generating accurate, TV-anime-style character reproductions.
Key Workflow Discoveries
During my testing, I focused on system adjustments and exploring Anima's strong ability to comprehend natural language. Here is what I found:
Natural Language Reduce Concept Bleeding:
Rewriting my prompts completely into natural language improved the generation success rate and made the blending between the character and the background much more natural. LLM-generated natural language prompts simply work better than relying on tag weights.Batch Generation via Wildcards:
Because Anima comprehends natural language so well, generating complex scenes is much smoother. When consistency is not an issue, you can simply ask an LLM (e.g.: Gemini or GPT) to write a series of scenarios describing a story, and then embed them into wildcards to batch process different actions.Improved Scene Recognition:
Scene recognition accuracy has noticeably improved for specific environments like "inside a Ferris wheel" or an "old bus stop". The model also properly generates broader concepts like a "sea of clouds". I also find it is good at rendering some famous landmark spots in the real world.Qwen's High Sensitivity:
The model handling prompt parsing (Qwen) is highly sensitive to syntax. The presence of a period, the position of newlines, and formal English sentence structures have a massive impact on the final result.High-Fidelity in Small-Scale Framing:
I observed that Anima is exceptionally good at rendering and framing a subject within a very small area of the canvas without destroying their overall shape or character similarity. This is incredibly useful when generating massive, expansive backgrounds where the character is just a small figure in the overall composition. It also holds up flawlessly when confining the subject to tight, localized areas, such as a reflection in a mirror, a glimpse through a window, or even featured on a poster.
Character Testing Results Without LoRAs
Anima-base v1.0’s training data goes up to September 2025. Depending on the character, you may not need a LoRA at all.
Works without LoRA:
Yakishio Lemon, Yaotome Urushi, Takagi-san, Suletta Mercury, Amy (sui sei no gargantia), Kofune MioFails without LoRA:
Higa Kana, Umeyashiki Sakurako, Yokoi Rumi
Additionally, these concepts do not require a LoRA nor ControlNet:
Shaving armpits in front of a mirror
shaving armpits with a razor
White vignette
Interacting with a silhouette subject
Core Guide: Prompting for the TV Anime Version
This section will verify the reproducibility of anime-version characters in Anima-Base 1.0. Here is the step-by-step progression to format your prompts properly.
1. Basic Structure for Popular Characters
Popular characters from before September 2015 can be generated with a straightforward structure.
Because it is highly sensitive to periods, line breaks, and formal English sentence structures, it is recommended to use as well-structured a syntax as possible.
You can optionally add "anime coloring".
[quality tags..., global styles...].
1girl, suletta mercury, gundam suisei no majo,
[character traits..],...2. Handling Minor Characters and Long Prompts
Specifying
@artistis necessary when the overall prompt is too long or when generating minor characters.You can search for character names, series titles, and artists on Danbooru.
This method is only effective when the original art style is close to the anime version's art style.
[quality tags..., global styles...].
1girl, suletta mercury, gundam suisei no majo, (@artist name: 1.0),
[character traits..],...3. Weight Adjustment for @artist
The influence of
@artistcan sometimes be as strong as a LoRA.If the art style becomes too rigid to change, please adjust it by lowering the weight.
[quality tags..., global styles...].
1girl, suletta mercury, gundam suisei no majo, (@artist name: 0.4),
[character traits..],...4. Syntax for Anime Version Reproduction
Ensure that "anime screenshot", the character name, and the series title are written in the same paragraph and placed as close together as possible.
Must put in the beginning of the paragraph.
Also, be sure to use commas (
,) instead of periods (.) as separators.
[quality tags..., global styles...].
anime screenshot, 1girl, suletta mercury, gundam suisei no majo, (@artist name: 0.4),
[character traits..],...5. Weight Adjustment for the "Anime Look"
If the overall prompt is too long, you might need to increase the weight of "anime screenshot".
[quality tags..., global styles...].
(anime screenshot: 2.0), 1girl, suletta mercury, gundam suisei no majo, (@artist name: 0.4),
[character traits..], ...If you still cannot reproduce the character's look after trying these methods, your only option is to use a LoRA. I hope this is helpful to everyone.
Appendix: My Prompt Header
The following are the core prompts I recently use. I built my prompt repository on top of a highly modularized architecture, which currently features 54 backgrounds, 6 characters, and 3 global anime styles. I find Anima handles all of them very consistently. Some style prompts are dependent on the subject, and the weight is also variable. An example header looks like this:
masterpiece, best quality, score_7, highres, absurdres, official art. A high-budget anime screenshot with fully rendered anime coloring, the subject is drawn with a 0.3mm manga G-pen, featuring distinctly delicate, exceptionally soft, color-matched strokes tracing the edges, mimicking the "color trace" effect on a fully colored key animation frame.
(anime screenshot: 2.0), anime coloring, (brown outlines, brown edges: 2.0), 1girl, suletta mercury, gundam suisei no majo, (@mogumo: 0.4), ... [character traits..].
[actions..].
[backgrounds..].
[lighting..].
[finishes..].Appendix: The Mentioned 6 Anime Characters
If you are interested, a combined thumbnail featuring the tachi-e of the 6 mentioned anime characters can be seen on my Pixiv post. Please note that the actual post does not include the character-only tachi-e used to create that thumbnail. Instead, the gallery contains the complete generations for each character, fully integrated with rendered backgrounds.
You will require an account to view them. While this specific post is completely SFW (https://www.pixiv.net/artworks/145974375) in the anime category, please be aware that my general Pixiv profile does contain R-18 content, so please browse accordingly based on your account settings.