Anima#
Anima is a compact anime and illustration family built on a 2-billion-parameter Cosmos-Predict2 base, seven checkpoints in total. It sits between SDXL and Z-Image in how you talk to it: quality tags and working negative prompts like SDXL, but it reads descriptive sentences better than tag lists.
Three sit under the Anima family entry in the model picker. The other four are direct picks.
At a glance#
| Model | Picker entry | Steps | CFG | Sampler | Scheduler | Pick it for |
|---|---|---|---|---|---|---|
| Anima Turbo | Anima > Turbo | 12 | 1 | dpmpp_2m_sde |
simple |
The fast one. Preview 3 with a turbo LoRA baked in. |
| Anima (Preview 3) | Anima > Preview 3 | 30 | 5 | dpmpp_2m_sde |
simple |
The original preview release. General anime and illustration. |
| Anima Base (v1.0) | Anima > Base 1.0 | 50 | 5 | euler_ancestral |
normal |
The first full release. Slowest and cleanest. |
| Anima (aesthetic-v1.1) | direct pick | 30 | 4 | not pinned | not pinned | An aesthetic tune. Smoother style, fewer artifacts. |
| Anima Cat Tower (v1.0) | direct pick | 30 | 5 | dpmpp_2m_sde |
simple |
A Cat Tower flavoured finetune with its own tag set. |
| AnimaYume (v0.5) | direct pick | 50 | 5 | euler_ancestral |
normal |
A Yume finetune of Base v1.0, same 50-step recipe. |
| WAI Anima (v1.0) | direct pick | 30 | 5 | euler_ancestral |
normal |
The WAI finetune, and the only one shipping its own hi-res fix. |
Every model here renders at 896x1152 by default, defaults to a batch of 2, and falls back to the global AnimeSharp (4x) upscaler unless you pick another. None of them pin their own default upscaler.
Steps and CFG are clamped, not rejected
Pass a value outside a model's recommended range and you get an image at the nearest legal value rather than an error. See Parameters.
Prompt style#
Anima understands both natural language and Danbooru-style tags, and it does better with natural language. Two or three descriptive sentences beat a tag list on all seven checkpoints. Tags still work, they just leave more of the frame to the model.
/imagine prompt: A girl with long white hair and a red ribbon stands on a rooftop at dusk, city lights blurred behind her, wind lifting her coat, warm rim light from the left model: Anima Base (v1.0)
Eimi adds quality tags for you on every model, so do not write them yourself. The exact strings differ per checkpoint and they materially change what you get:
| Model | Added to your positive | Added to your negative |
|---|---|---|
| Anima Turbo, Anima (Preview 3), Anima Base (v1.0), AnimaYume (v0.5) | masterpiece, best quality, score_7 |
worst quality, low quality, score_1, score_2, score_3, artist name |
| Anima (aesthetic-v1.1) | masterpiece, best quality, score_7, smoother style, reduced artifacts |
nothing |
| Anima Cat Tower (v1.0) | masterpiece, best quality, highres, absurdres |
worst quality, low quality, blurry, jpeg artifacts, lowres |
| WAI Anima (v1.0) | masterpiece, best quality, score_7 |
worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, lowres, censor |
The artist name negative on five of the seven is worth knowing about: it actively pushes against artist-style prompting. Cat Tower drops it and adds resolution tags instead, which is why it comes back sharper and busier than the rest of the family on the same prompt.
With your NSFW setting off, every Anima model except aesthetic-v1.1 also adds nsfw, explicit to the negative and (safe:1.2) to the positive. aesthetic-v1.1 ships no safe-content guard strings at all, so that toggle changes nothing on it.
Negative prompts work here. Unlike Z-Image or Flux, what you type in negative: is honoured, and at the family's default CFG of 4 to 5 it has real pull. Anima Turbo is the exception -- it runs at CFG 1, where a negative barely registers.
Six of the seven (everything except aesthetic-v1.1) carry a model-specific instruction set for the AI enhancer, which writes flowing sentences rather than tags and deliberately leaves the quality tags alone. See AI Enhance and the prompting guide.
Character names come out as tags, not prose
Character matching formats detected names in escaped Danbooru form on this family -- hatsune_miku \(vocaloid\) -- even though the rest of the prompt is prose. That is a mismatch with how the enhancer writes, and it is worth knowing if a character render comes back looking confused.
The seven models#
Anima Turbo#
Anima Preview 3 with the Anima turbo LoRA baked in at full strength. Pick it when you want an Anima render now rather than in a minute.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | cosmos |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported, but weak at CFG 1 |
| Steps | 12 (8-16) |
| CFG | 1 (1-2) |
| Sampler | dpmpp_2m_sde |
| Scheduler | simple |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: quick anime and illustration drafts, iterating on a composition, and getting an Anima-looking image for roughly a quarter of the sampling work the 50-step models want.
Weak at: everything the turbo LoRA costs. At CFG 1 the negative prompt and the worst quality, low quality, score_1... list Eimi appends for you are both close to inert, so you lose the quality floor the rest of the family gets for free. Seed variety is narrower than Preview 3, which it is otherwise identical to.
Four of your own LoRAs will push the turbo LoRA out
Anima Turbo is Preview 3 plus an auto-added turbo LoRA, and it shares the family's 4-slot LoRA stack. The turbo LoRA is added after your picks, so if you select four LoRAs yourself it drops off the end and you silently get plain Preview 3 running a 12-step CFG-1 recipe it was never tuned for. Keep it to three if you want the turbo behaviour. The auto-added LoRA is hidden from the parameters view, so nothing on the result card will tell you this happened.
Anima (Preview 3)#
The original Anima preview release, and the checkpoint Anima Turbo is built from. A reasonable general-purpose pick.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | cosmos |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported and effective at CFG 5 |
| Steps | 30 (30-50) |
| CFG | 5 (4-6) |
| Sampler | dpmpp_2m_sde |
| Scheduler | simple |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: general anime and illustration, honouring negatives, and giving genuinely different compositions across seeds.
Weak at: being superseded. It is a preview build and Anima Base (v1.0) is the finished release of the same model, so unless you specifically prefer the preview's look there is a newer option. Its config also notes that er_sde is the sampler the author recommends, while the shipped default is dpmpp_2m_sde -- if results look muddy, try sampler: er_sde. At 30 steps and 896x1152 it is not fast, and 2B is a small model, so complex multi-character scenes fall apart sooner than they would on a bigger architecture.
Anima Base (v1.0)#
The first full Anima release, newer than Preview 3. The longest recipe in the family at 50 steps.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | cosmos |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported and effective at CFG 5 |
| Steps | 50 (30-60) |
| CFG | 5 (4-6) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: the cleanest output in the family, strong response to both positive and negative prompts, and the widest step range (30 to 60) if you want to trade time for finish.
Weak at: speed, plainly. At 50 steps and batch 2 it is over four times the sampling work of Anima Turbo per card, and this is one GPU in somebody's house, so a busy queue makes that wait obvious. euler_ancestral also injects fresh noise at every step, which means the same seed at 40 steps and at 50 steps gives you two different pictures -- a nuisance if you are trying to tune a render you already liked. It ships no style of its own, so bare prompts come back flatter than the finetunes give you.
Anima (aesthetic-v1.1)#
An aesthetic tune of the 2B base whose baked-in tags ask directly for a smoother style and fewer artifacts. The odd one out in the family's configuration.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | none |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported, but the model adds nothing of its own |
| Steps | 30 (not set) |
| CFG | 4 (not set) |
| Sampler | not pinned |
| Scheduler | not pinned |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: smooth, low-artifact illustration without you having to ask for it, and being the one Anima model where your nsfw setting does not silently rewrite your prompt in either direction.
Weak at: predictability. It pins no sampler, no scheduler and no recommended ranges, so /imagine's generic defaults apply and your results will not match anyone else's unless you both pass sampler: and scheduler: explicitly. It ships no negative tags at all, so the quality floor the rest of the family gets from worst quality, low quality, score_1... is missing -- write your own negative or expect more junk frames. It is also the only Anima model with no custom enhancer instructions, so AI enhancement produces generic prose here rather than Anima-shaped prose.
This entry is under-specified on purpose or by accident, and it shows
Steps 30 and CFG 4 are all it sets. No sampler, no scheduler, no recommended minimum or maximum. Treat any result from it as less reproducible than the rest of the family.
Anima Cat Tower (v1.0)#
A Cat Tower flavoured finetune. Its baked-in tag set is the most different in the family and that is the reason to pick it.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | cosmos |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported and effective at CFG 5 |
| Steps | 30 (30-50) |
| CFG | 5 (4-6) |
| Sampler | dpmpp_2m_sde |
| Scheduler | simple |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: sharp, detail-dense illustration. Its positive tags are masterpiece, best quality, highres, absurdres rather than the family's score_7 line, and it is the only Anima model with a real negative list that does not include artist name, so artist-style prompting has a chance of landing here.
Weak at: clean flat areas and simple compositions. highres, absurdres pushes toward busy, over-detailed frames whether you wanted that or not, and at 896x1152 the model does not have the pixels to pay for the detail it is asking for -- fine patterns turn to noise. Its picker description is also stale: it still calls the model v0.5 when the checkpoint is v1.0, and it recommends er_sde while the model actually ships dpmpp_2m_sde as its default.
AnimaYume (v0.5)#
A Yume finetune built on Anima Base v1.0, running the same 50-step recipe.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | cosmos |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported and effective at CFG 5 |
| Steps | 50 (30-60) |
| CFG | 5 (4-6) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: the Anima Base look with a different aesthetic lean. Same tag set, same sampler, same ranges.
Weak at: the same things as its parent, and for the same reasons: it is Anima Base v1.0 with different training and an identical recipe, so 50 steps of sampling, euler_ancestral seed instability, and 2B's ceiling on complex scenes all carry over unchanged. Being a v0.5, it is also an earlier point in its own line than Base v1.0 is in its, so it is the less finished of the two.
WAI Anima (v1.0)#
The WAI finetune, and the only Anima model that ships a hi-res fix recipe of its own.
| Architecture | Anima 2B (Cosmos-Predict2) |
| Subtype | cosmos |
| Prompting | Either; prose beats tags. Two to three descriptive sentences. |
| Negatives | Supported and effective at CFG 5 |
| Steps | 30 (25-40) |
| CFG | 5 (4-6) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | none pinned, falls back to AnimeSharp (4x) |
Good at: finishing. Its author-supplied hi-res fix recipe (1.5x, 20 steps, denoise 0.4, R-ESRGAN 4x+ Anime6B) is tuned for this checkpoint, so the second pass is more likely to sharpen than to redraw. Its extra negatives -- blurry, jpeg artifacts, lowres, censor on top of the family list -- give it the strongest built-in quality floor here.
Weak at: subtlety, as a consequence. That long baked-in negative is appended to whatever you type, so it is harder to get a deliberately soft, grainy or low-fi look out of this model than out of the others. Its recommended step range is the narrowest in the family (25 to 40), so there is less room to trade time for quality. euler_ancestral again means the same seed will not survive a step-count change.
How Anima differs from SDXL#
Same quality-tag and negative-prompt mechanics, different machine underneath.
| Anima | SDXL family | |
|---|---|---|
| Base resolution | 896x1152 for every model | 1152x1536 on Illustrious, 896x1152 on NoobAI / Pony |
| Optimal aspect ratios | 5 (1:1, 9:7, 7:9, 4:3, 3:4) | 15 to 17 depending on subtype |
| Typical step count | 30 to 50 | 12 to 35 |
| Prompt shape | Tags accepted, prose preferred | Tags, strongly |
| Model loading | Diffusion model, text encoder and VAE loaded separately | One bundled checkpoint |
| LoRA pool | 7 Anima-compatible LoRAs in the library | 28 SDXL-compatible LoRAs |
| Model size | 2 billion parameters, anime focused | SDXL class |
What is the same: quality tags, negative prompts, 4 LoRA slots affecting both the model and the text encoder, Detailer, hi-res fix, and every upscale method. Anima is one of the families that does accept the Detailer, so faces, eyes and hands can be repaired after the fact -- something Z-Image cannot do.
Anima ratios outside the five optimal ones are computed, not tuned
Ask for 16:9 or 21:9 on an Anima model and the size is derived from the model's pixel budget rather than looked up in a tuned table. Composition can drift. Stick to the five listed ratios when you can.
The 7-LoRA pool is the family's real constraint. Whatever you want to restyle an Anima render with probably has not been trained for it, and an SDXL or Z-Image LoRA will not load on this architecture. See the LoRA catalog.
Resolutions#
Default 896x1152 (3:4), shared by all seven models.
| 1:1 | 9:7 | 7:9 | 4:3 | 3:4 |
|---|---|---|---|---|
| 1024x1024 | 1152x896 | 896x1152 | 1152x864 | 864x1152 |
A literal WIDTHxHEIGHT is accepted too, snapped down to a multiple of 8 and floored at 64. Anima tops out around 1 megapixel natively, which is well under what Krea 2 or Z-Image start at -- if you want a large final image from an Anima model, render at the native size and get there with hi-res fix or /upscale.
Hi-res fix#
WAI Anima (v1.0) is the only model in the family that ships its own hi-res fix recipe: upscale factor 1.5, 20 steps, denoise 0.4, using the R-ESRGAN 4x+ Anime6B upscaler. That upscaler is in the catalog specifically because the model's author recommends it for this pass.
Every other Anima model falls back to the global defaults: denoise 0.35, second-pass steps at 75% of the parent job's, and the AnimeSharp (4x) upscaler. Those are generic numbers, not Anima numbers, so expect to adjust the denoise down if a second pass is rewriting faces. See Hi-Res Fix and Detailer and the upscaler catalog.
Related#
- Model overview -- the full catalog
- SDXL family -- the closest relative in prompt mechanics
- /imagine -- every parameter
- LoRA catalog