Testing Official Anima Turbo v1.1 and Aesthetic v1.1 on M1 Max to See If Base-Trained Character LoRAs Work
Contents
Anima Turbo v1.1 appeared on the official Civitai page.
I normally train and run character LoRAs (small adapter weights designed to generate specific characters) on Anima-Base v1.0, but the same page lists official variants: Anima-Aesthetic and Anima-Turbo.
The README states that “LoRA should be trained on Base.” Aesthetic (refined style and reduced artifacts) and Turbo (distilled for fast, few-step generation) are derivatives created from Base.
The biggest challenge with Anima is whether character LoRAs actually render the intended character.
Back in May, in my comparison of Anima-Base v1.0 and WAI-Anima, applying a LoRA to a different base caused character breakdown depending on the seed.
I tested whether character identity is preserved when applying a Base-trained LoRA to Aesthetic or Turbo.
I tested 5 model configurations on an M1 Max 64GB using ComfyUI, keeping prompt, seed, and resolution identical while running each model family at its standard step count (8 steps for Turbo, 30 steps for Base).
Basic Information
| Item | Details |
|---|---|
| Authors | CircleStone Labs, Comfy Org |
| Base Model | NVIDIA Cosmos-Predict2-2B-Text2Image |
| Parameter Count | Approx. 2 Billion |
| File Size | Aesthetic and Turbo: 4,182,230,656 bytes each; Base: 4,182,218,328 bytes (approx. 4.18 GB each) |
| Text Encoder / VAE | Qwen3 0.6B / Qwen-Image VAE (Shared with Base) |
| License | CircleStone Labs Non-Commercial License |
| Download | Hugging Face, Civitai |
The official models are split into three main families.
| Family | README Description | Releases and Dates (UTC) |
|---|---|---|
| Anima-Base | Pre-trained model. Highest freedom, diversity, and style compliance. Train LoRAs on this | base-v1.0 (May 14, 2026) |
| Anima-Aesthetic | Fine-tuned to enhance output consistency and default artistic quality | v1.0 (July 8), v1.0b (July 9), v1.1 (July 13) |
| Anima-Turbo | Distilled for few-step generation. Stable default style, but lower diversity | v1.0 (July 8), v1.1 (August 24) |
The v1.1 changes are briefly described on Civitai.
| Version | Civitai Description |
|---|---|
| Turbo v1.1 | Retrained distillation and style adjustments. Less flat, more detail, and better adherence to artist styles |
| Aesthetic v1.1 | Retrained to adjust style. Smoother, fewer artifacts, while preserving fine details |
The recommended settings from the README are as follows:
| Item | Base / Aesthetic | Turbo |
|---|---|---|
| Resolution | 512² to 1536² | 512² to 1536² |
| Steps | 30 to 50 | 8 to 12 |
| CFG | 4 to 5 | 1 |
| Sampler | er_sde, euler_a, dpmpp_2m_sde_gpu, euler | euler recommended |
| Positive Prefix | masterpiece, best quality, score_7, safe, | masterpiece, best quality, score_7, safe, |
| Negative | worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration | Same as left |
| Score Tags | Recommended not to use for Aesthetic | Supported |
Compared Models
| Label | Model | Role |
|---|---|---|
| Base + Turbo LoRA | anima-base-v1.0 + anima-turbo-lora-v0.1 (strength 1.0) | Standard setup from the Anima-Base v1.0 article. Baseline for Turbo |
| Turbo v1.0 | anima-turbo-v1.0 | Initial release of official Turbo |
| Turbo v1.1 | anima-turbo-v1.1 | The updated version |
| Base | anima-base-v1.0 | Baseline for 30-step generation. Training source for the LoRA |
| Aesthetic v1.1 | anima-aesthetic-v1.1 | The updated version |
The 3 Turbo models ran at 8 steps, while Base and Aesthetic v1.1 ran at 30 steps.
Generation Conditions
| Parameter | 3 Turbo Models | Base / Aesthetic v1.1 |
|---|---|---|
| Resolution | 832×1216 (832×1024 for portraits, 1216×832 for two characters) | Same as left |
| Steps | 8 | 30 |
| CFG | 1.0 | 4.5 |
| Sampler / Scheduler | er_sde / simple | Same as left |
| Seed | 42 | Same as left |
| Character LoRA | keikana-animabase-v2_epoch140 (strength_model 1.0, strength_clip 1.0) | Same as left |
| Positive Prefix | masterpiece, best quality, safe | Same as left |
| Environment | M1 Max 64GB, ComfyUI (127.0.0.1:8188) | Same as left |
The character LoRA is the Kei-Kana dual-character LoRA v2 trained on Anima-Base v1.0.
I used the same files and prompts as in the June AMIXA comparison.
Per Aesthetic’s README recommendations, score tags were omitted across all models for consistency.
The sampler was fixed to er_sde to align with past comparison articles.
Base Style Without LoRA
I generated images without any character LoRA using a standard blonde hair and white robe prompt.
Test 1: Full Body on White Background
| Model | Observations |
|---|---|
| Base + Turbo LoRA | Gold embroidery and a light blue apron; highest color variety among the 3 Turbo variants |
| Turbo v1.0 | White and muted gold only; overall colors are pale |
| Turbo v1.1 | Similar palette to v1.0, but with finer embroidery and deeper shadows |
| Base | Bright gold fills the apron and hem; blunt bangs with flatter shading |
| Aesthetic v1.1 | Exposed forehead, muted gold, long trailing hem; strongest three-dimensional shading |
The two dedicated Turbo models kept their outfit palettes restrained. Turbo v1.1 increased embroidery and shadow details compared to v1.0, which aligns with Civitai’s changelog.
The two 30-step models diverged in facial features and hair shape. Base leaned toward flat coloring, while Aesthetic v1.1 showed a distinct, muted palette with pronounced depth.
Prompt
masterpiece, best quality, safe, 1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, standing, looking at viewer, full body, white backgroundNegative:
worst quality, low quality, artist name, blurry, jpeg artifacts, chromatic aberrationTest 2: Dynamic Running Scene in Meadow
| Model | Observations |
|---|---|
| Base + Turbo LoRA | Large castle on the lower left, sunset clouds, foreground grass |
| Turbo v1.0 | Castle is small and distant; character centered in the frame |
| Turbo v1.1 | Castle placement and scale resemble Base + Turbo LoRA |
| Base | Cape lining turns red; castle remains small and distant |
| Aesthetic v1.1 | Draws the hilltop castle, distant mountain ridges, and foreground rocks |
The 3 Turbo models shared similar composition; the main differences were the size and distance of the castle.
The 30-step models positioned the character slightly further back, with Aesthetic v1.1 rendering the highest amount of background detail.
Prompt
masterpiece, best quality, safe, 1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, running, wind, hair blowing, dynamic pose, fantasy landscape, castle in background, sunset sky, dramatic clouds, grass fieldNegative is identical to Test 1.
Applying Character LoRA
Next, I applied the character LoRA trained on Anima-Base.
Kana-chan’s identifying features are brown hair, an ahoge, a right-side ponytail, a blue scrunchie, brown eyes, and a white shirt with a red tie.
Kei-chan’s features are long blonde hair, blunt bangs, side braids, a blue ribbon behind the head, blue eyes, and a red ribbon tie.
Kana-chan
Full-body outputs:
For portrait shots, the prompt changed to upper body, portrait:
| Model | Observations |
|---|---|
| Base + Turbo LoRA | All features appear; black tights in full-body, tie becomes a large red ribbon in portrait |
| Turbo v1.0 | All features appear; bare legs with short socks in full-body |
| Turbo v1.1 | All features appear; knee-high socks in full-body |
| Base | All features appear; black tights in full-body, navy blazer and red ribbon in portrait |
| Aesthetic v1.1 | All features appear; hands clasped in front in full-body, navy blazer and red ribbon in portrait |
Kana-chan’s key traits—brown hair, ahoge, side ponytail, and blue scrunchie—were rendered accurately across all 5 models.
Differences were limited to legwear, tie shape, and arm positioning, with no distortion in hair or face structure.
Kei-chan
Full-body outputs:
Portrait outputs:
| Model | Observations |
|---|---|
| Base + Turbo LoRA | Hair traits and red ribbon tie both match |
| Turbo v1.0 | Hair traits appear, but neck ribbon turns navy blue |
| Turbo v1.1 | Hair traits appear, but neck ribbon turns navy blue; jacket also turns navy in full body |
| Base | Hair traits and red ribbon tie match; plaid skirt |
| Aesthetic v1.1 | Hair traits and red ribbon tie match; hands clasped in front in full body |
Kei-chan’s hairstyle structure remained intact across all 5 models.
The only divergence was the neck ribbon color: both dedicated Turbo models generated navy blue ribbons on seed 42, whereas Base + Turbo LoRA, Base, and Aesthetic v1.1 rendered them red. The prompt token blue ribbon (intended for the rear hair ribbon) likely influenced the neck ribbon.
Prompt (Kana & Kei)
Kana full-body:
masterpiece, best quality, safe, 1girl, solo, kanachan, brown hair, side ponytail, ahoge, brown eyes, school uniform, full body, standing, looking at viewer, white background, simple backgroundKei full-body:
masterpiece, best quality, safe, 1girl, solo, keichan, blonde hair, long hair, blunt bangs, blue ribbon, blue eyes, school uniform, full body, standing, looking at viewer, white background, simple backgroundFor portraits, full body, standing was replaced by upper body, portrait, and full body, knees, feet, lower body was added to negative prompts.
Negative:
worst quality, low quality, blurry, lowres, jpeg artifacts, bad hands, bad anatomy, fused fingers, fused legs, fused limbs, extra arms, extra legs, conjoined twins, deformed faceTwo-Character Outputs
Without physical contact, both characters stood side-by-side (standing side by side, full body). Per README recommendations, left-right positioning and detailed hair features were included in the prompt.
| Model | Observations |
|---|---|
| Base + Turbo LoRA | Clean separation (Kei left, Kana right), no ahoge bleeding; Kei has red ribbon |
| Turbo v1.0 | Clean separation and ahoge intact; Kei’s ribbon becomes a thin navy ribbon |
| Turbo v1.1 | Clean separation and ahoge intact; Kei has red ribbon |
| Base | Clean separation and ahoge intact; Kei has red ribbon, Kana wears black tights |
| Aesthetic v1.1 | Clean separation and ahoge intact; Kei has red ribbon; both clasp hands in front |
Character separation held up cleanly without feature bleeding across all models. Only Turbo v1.0 shifted Kei’s neck ribbon to navy.
Prompt
masterpiece, best quality, safe, 2girls, multiple girls, two different girls, kanachan, keichan, standing side by side, full body, looking at viewer, smile, school uniform, keichan on the left has long blonde hair with blunt bangs and a blue ribbon, kanachan on the right has brown hair with a side ponytail and an ahoge, white background, simple backgroundNegative added 1girl, solo to the standard negative list.
Measuring LoRA Reliability Across 4 Seeds
Testing only a single image on seed 42 makes it hard to distinguish model consistency from random chance.
To assess consistency, I generated 4 images per configuration using seeds 42, 100, 200, and 300, and counted how often the characters emerged correctly.
For modifications to outfits and hairstyles, I checked 2 seeds (42 and 100).
| Test Condition | Base + Turbo LoRA | Turbo v1.0 | Turbo v1.1 | Base | Aesthetic v1.1 |
|---|---|---|---|---|---|
| Kana (Trigger only: brown hair, ahoge, side ponytail) | 0/4 | 0/4 | 0/4 | 0/4 | 0/4 |
| Kei (Trigger only: blonde long, blunt bangs, braid, blue ribbon) | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 |
| Kana (Detailed features: brown hair, ahoge, side pony, blue scrunchie, red tie) | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 |
| Kei (Detailed features: blonde, blunt bangs, braid, blue ribbon) | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 |
| Kei (Red ribbon tie at neck) | 4/4 | 1/4 | 3/4 | 4/4 | 4/4 |
| 2 Characters (Count: exactly 2 girls) | 4/4 | 4/4 | 4/4 | 4/4 | 1/4 |
| 2 Characters (Separation: clean left/right, no ahoge bleeding) | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 |
| 2 Characters Kei neck ribbon (Red ribbon tie) | 4/4 | 1/4 | 4/4 | 4/4 | 4/4 |
| Kana outfit change (Wears specified clothes, retains features) | 2/2 | 2/2 | 2/2 | 2/2 | 2/2 |
| Kei hairstyle change (Becomes twintails, retains features) | 1/2 | 2/2 | 2/2 | 1/2 | 1/2 |
I inspected every image individually to verify the criteria.
For Aesthetic v1.1 in the two-character test, even when extra characters appeared, each individual person was identifiably either Kei or Kana, so character separation scored 4/4 while character count was tracked separately.
Trigger Words Only
Using only kanachan or keichan with basic body and background tags.
Kana-chan:
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
Kei-chan:
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
For Kana, 0 out of 4 images came out correctly across all 5 models.
Instead, the output depicted a character with Kei’s hairstyle: long hair with side braids and a rear ribbon, with hair colors alternating between brown and blonde.
Because this occurred identically across all 5 models, it is a trait of this merged LoRA rather than a checkpoint discrepancy. Captions during training likely required explicit hairstyle tags to activate Kana properly.
Conversely, Kei’s hairstyle appeared across all 20 images using the trigger word alone.
However, without clothing prompts, outfits varied widely: bodysuits, white dresses, military jackets, swimwear, and chef coats.
On seed 100 for Base and Aesthetic v1.1, the output rendered naked bodies despite safe in the positive prompt. Aesthetic v1.1 produced underwear or bare-apron outputs in 3 out of 4 images even with Kana’s trigger. The 3 Turbo models produced 0 nude outputs across 24 trigger-only images.
With Explicit Character Features
I ran 4 seeds (42, 100, 200, 300) with explicit hair and eye feature tags, identical to the previous section.
Kana-chan:
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
Kei-chan:
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
Two-Character:
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
When hair and eyes were explicitly specified, both Kana and Kei appeared accurately across all 5 models.
Aesthetic v1.1 frequently generated characters with clasped hands in front of the body (4/4 for Kana, 3/4 for Kei), despite no such instruction in the prompt.
In two-character prompts, the 3 Turbo models and Base maintained left/right separation across all 16 images without ahoge contamination.
Aesthetic v1.1 spawned additional people in 3 out of 4 seeds: seeds 100 and 200 generated 4 people (two Keis and two Kanas), while seed 300 generated 3 people (two Keis and one Kana). The individual characters did not blend features. Because Base produced 0 duplicates under identical prompts, LoRA, and step counts, the additional fine-tuning in Aesthetic v1.1 appears to weaken numeric count adherence.
The primary point of divergence was Kei’s neck ribbon: Base + Turbo LoRA was 4/4 red; Turbo v1.0 was only 1/4 red (3/4 navy); Turbo v1.1 improved to 3/4 red in solo and 4/4 in two-character outputs. Base and Aesthetic v1.1 stayed 4/4 red across both tests.
Testing Prompt Adaptability
I checked whether character identity was preserved when intentionally altering clothing or hairstyle.
Outfit Change (Kana-chan)
Specified a black hoodie, denim shorts, and sneakers.
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
Hairstyle Change (Kei-chan)
Specified twintails.
Base + Turbo LoRA
Turbo v1.0
Turbo v1.1
Base
Aesthetic v1.1
The outfit instruction succeeded in all 10 images across all 5 models while retaining brown hair, the ahoge, the side ponytail, and the blue scrunchie.
For the hairstyle modification, models utilizing Base weights (Base + Turbo LoRA, Base, and Aesthetic v1.1) failed on seed 100 and remained in her default long hairstyle. The two dedicated Turbo models followed the prompt on both seeds. On Turbo v1.0 seed 42, the result leaned toward a two-side-up tied high, but satisfied the twintail criteria.
Prompt (Triggers & Changes)
Kana (Trigger only; for Kei, replace kanachan with keichan):
masterpiece, best quality, safe, 1girl, solo, kanachan, full body, standing, looking at viewer, white background, simple backgroundKana outfit change:
masterpiece, best quality, safe, 1girl, solo, kanachan, brown hair, side ponytail, ahoge, brown eyes, black hoodie, denim shorts, sneakers, full body, standing, looking at viewer, white background, simple backgroundKei hairstyle change:
masterpiece, best quality, safe, 1girl, solo, keichan, blonde hair, twintails, blue eyes, school uniform, full body, standing, looking at viewer, white background, simple backgroundNegative is identical to previous sections.
Background Density
Kana-chan seated by a cafe window.
| Model | Observations |
|---|---|
| Base + Turbo LoRA | Renders window, plants, bookshelves, teacup, ceiling lamps, and rear chairs |
| Turbo v1.0 | Close-up with chin in hand; window flowers and bookshelf |
| Turbo v1.1 | Close-up with chin in hand; vest removed, white shirt + red ribbon |
| Base | Frontal framing holding cup with both hands; two ceiling lamps, curtains, rear chairs |
| Aesthetic v1.1 | Frontal framing holding cup; shelf teapot, wall lamp, tall houseplant |
The dedicated Turbo models placed the character closer to the camera and rendered fewer background elements.
The two 30-step models matched the frontal composition and drew background depth on par with Base + Turbo LoRA.
Prompt
masterpiece, best quality, safe, 1girl, solo, kanachan, brown hair, side ponytail, ahoge, brown eyes, school uniform, sitting at a cafe table by the window, potted plants, bookshelf, teacup, warm afternoon sunlight, detailed background, intricate details, looking at viewerNegative:
worst quality, low quality, blurry, lowres, jpeg artifacts, bad hands, bad anatomy, deformed faceImage-to-Image (i2i)
Using the same source image generated by WAI-Illustrious v160 with the Kana-chan LoRA from my April WAI-Anima article, I ran i2i at denoise 0.5 without any character LoRA applied.
| Model | Observations |
|---|---|
| Base + Turbo LoRA | Preserves composition, gold circle, and sash knot; floral pattern inside the circle |
| Turbo v1.0 | Circle remains plain; closest reproduction of the source image |
| Turbo v1.1 | Circle gains patterns and distorted outline |
| Base | Preserves composition, circle, and sash knot; floral pattern inside the circle |
| Aesthetic v1.1 | Very similar to Base; circle pattern is slightly fainter |
All 5 models rendered rounder faces; the character looked younger than the source, similar to what I observed in April.
Prompt
masterpiece, best quality, safe, 1girl, solo, brown hair, side ponytail, ahoge, brown eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, standing, looking at viewer, full bodyNegative is identical to the base style test.
Generation Speed
Measured on the same M1 Max 64GB machine using ComfyUI execution logs.
Test 1 is the first run including initial model loading; Test 2 is a subsequent run at the same resolution (832×1216).
| Model | Steps | Test 1 | Test 2 | Portrait (832×1024) |
|---|---|---|---|---|
| Base + Turbo LoRA | 8 | 42.3s | 38.1s | 30.2s |
| Turbo v1.0 | 8 | 40.2s | 38.1s | 30.2s |
| Turbo v1.1 | 8 | 40.2s | 38.2s | 30.2s |
| Base | 30 | 262.8s | 258.7s | 210.7s |
| Aesthetic v1.1 | 30 | 261.2s | 259.1s | 208.9s |
Switching to a standalone Turbo model yielded nearly identical generation times compared to applying the Turbo LoRA on Base.
In my May Anima-Base v1.0 benchmark, Base + Turbo LoRA took 46–48 seconds; the improvement here likely reflects ComfyUI performance optimizations.
The two 30-step models took roughly 260 seconds per image, roughly 6.8 times longer than the Turbo models.
Because CFG 1 skips negative evaluation, model passes total 8 versus 60 (30 steps × 2 evaluations for CFG 4.5), which closely matches the measured timing ratio. Base and Aesthetic v1.1 had identical execution times.
Recommended Use Cases
Based on these tests, here is how each variant fits practical workflows when applying Base-trained character LoRAs:
| Use Case | Recommended Model | Evidence |
|---|---|---|
| Fast character generation | Base + Turbo LoRA, Turbo v1.1 | ~38s per image. 4/4 accuracy when specifying hair and eyes. Red neck ribbon appeared in 8/8 for Base + Turbo LoRA, 7/8 for Turbo v1.1 |
| Exact outfit color adherence | Base + Turbo LoRA, Base | Kei’s red ribbon tie maintained 4/4 in solo and two-character tests. Turbo v1.0 dropped to 1/4 |
| Multi-character scenes | Base + Turbo LoRA, Turbo v1.1, Base | Count and separation scored 4/4. Aesthetic v1.1 spawned 3–4 characters in 3 out of 4 seeds |
| Modifying hairstyles | Turbo v1.0, Turbo v1.1 | Twintails followed in 2/2. Base-weight models stayed long on seed 100 |
| Detailed backgrounds | Base + Turbo LoRA, Base, Aesthetic v1.1 | Renders deep room elements like lamps and background furniture. Dedicated Turbo zooms in |
| Default artistic styling without LoRA | Aesthetic v1.1 | Muted palette and solid depth with rich background detail (~260s/image) |
| Preserving source in i2i | Turbo v1.0 | Gold circle stays plain, closest to source image |
NSFW Generation Test
Images in this section contain sensitive adult content and have been blurred for publication.
Generated without character LoRAs under two conditions: without the nsfw tag and with the nsfw tag (straight-on frontal framing). safe was removed from positive prompts.
Without nsfw Tag
With nsfw Tag (straight-on)
All 5 models generated nude figures regardless of whether the nsfw tag was included. Neither the standalone Turbo models nor Aesthetic v1.1 showed any output suppression.
Prompt
Without tag:
masterpiece, best quality, 1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, sitting on edge of spring, feet in water, forest, natural light, serene, full body, looking at viewerWith tag:
masterpiece, best quality, 1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white backgroundNegative is identical to the base style test.