Seven 2B Anima Derivatives With a Base-Trained Character LoRA, and Only RDBT Kept Kei's Ribbon Tie Red in Both Full-Body and Bust Shots
Contents
I applied a character LoRA (a small add-on weight file that teaches a model specific characters) trained on Anima-Base v1.0 to the latest versions of seven 2B Anima derivatives, on an M1 Max 64GB in ComfyUI, with the same prompts and seed as Base.
Short version: all seven reproduced both characters’ hairstyles, but only RDBT v2.3 kept Kei’s red ribbon tie (the one at her neck) in both the full-body and bust shots, and RDBT also doubled two characters into four. Each result is a single image per model at seed 42.
Hotaru Blend and CapAnima finish an image in about 38 seconds at 8 steps, against roughly 260 seconds for Base at 30 steps.
Background: when I put together the October 2026 list of current Anima derivative checkpoints, many of the derivatives I tried in June had new versions.
Hikari, AnimaIka, AnimaYume, and RDBT moved to newer versions, and Nova Anime AM, CottonAnima, AMIXA, and JANIMA now have 2.9B versions.
My character LoRAs are trained on Anima-Base v1.0, so I wanted to know whether the characters still come out right on these newer derivatives.
I used the same prompts, seed, and LoRA as in the Turbo v1.1 and Aesthetic v1.1 comparison.
The four 2.9B models need a different ComfyUI version and a different way of applying the LoRA, so they are in a separate post on the four 2.9B models.
Models and Basic Information
The targets are the seven models in the list’s table of “current (2B) versions of derivatives from the June list.”
Civitai’s download API worked without login for only three of them (Hotaru, RDBT, and AnimaYume). The other four returned 401 without an API key.
I downloaded those four from Civitai in a browser and checked their SHA256 hashes against the values listed on Civitai.
| Model | Version | Release date (UTC) | Size | File | Author’s recommended settings |
|---|---|---|---|---|---|
| Hotaru Blend Anima | v6.0 | Sep 29 | 2B, Turbo LoRA baked in | bf16, 4.18GB | Euler A or ER SDE, Normal. 8–16 steps and CFG 1.0–1.5 for Turbo |
| CapAnima | Turbo 1.2 | Sep 25 | 2B, Turbo LoRA applied | bf16, 4.43GB | No numeric settings given |
| RDBT | v2.3 16s | Sep 19 | 2B, distilled | fp16, 4.18GB | Distilled versions: CFG 1, 8 or 16 steps. Always specify a style with @style |
| AnimaYume | v1.5 (base) | Sep 6 | 2B | bf16, 4.18GB | No numeric settings given. Complex prompts recommended |
| Anima Cat Tower | v1.1 | Sep 29 | 2B | bf16, 4.18GB | Specifies the order of a natural-language prompt. No sampler specified |
| AnimaIka | v4.9 | Oct 7 | 2B | 4.18GB | ER SDE or Res Multistep, Simple, 20–30 steps (28 recommended), CFG 2.5–5 (3.5 recommended) |
| Hikari Anima | v1.2.0 | Oct 16 | 2B | 4.18GB | 14–32 steps, ER SDE and others, Simple and others, CFG 2–6 (3.5 recommended), Flow Shift 2.5 |
For RDBT, I used v2.3 16s, which was the latest version on the checkpoint page when I ran the tests on October 21. On the 22nd, v3.0 base sma appeared on the same page.
These are the parts of each author’s description that relate to the results.
| Model | Author’s description |
|---|---|
| Hotaru Blend v6.0 | Rebuilt from scratch with a detailed, slightly cute, vivid style. Turbo LoRA baked in |
| CapAnima Turbo 1.2 | The author says they merged LoRAs after hearing that Anima is weak at chibi-style art. Turbo 1.2 is the Aesthetic version with a Turbo LoRA applied, and is the recommended one if you use Turbo |
| RDBT v2.3 16s | Trained to have no default style. Without @style, styles mix at random |
| AnimaYume v1.5 | Trained from Anima Base v1.0 for 12 epochs on 5 million images from Danbooru, e621, Gelbooru, and Konachan. LoRAs trained on Anima v1.0 or its derivatives may all break |
| Anima Cat Tower v1.1 | A fine-tune that pushes toward an anime look. For v1.1, the training captions were regenerated with Qwen 3.8/3.6 or Gemma 4 and all corrected by hand, to stabilize the style with natural-language prompts |
| AnimaIka v4.9 | Reduced instability on seeds that produce multiple characters. Slightly flatter |
| Hikari Anima v1.2.0 | A model built from Anima Base that does not need quality tags and avoids the glossy “AI masterpiece” look common in merges. v1.2.0 reduces broken compositions |
Baseline and Generation Settings
The baseline is Anima-Base v1.0 at 30 steps. I reused the images from the Turbo v1.1 and Aesthetic v1.1 comparison and did not regenerate them.
| Setting | Hotaru Blend v6.0, CapAnima Turbo 1.2 | RDBT v2.3 16s | AnimaYume v1.5, Cat Tower v1.1, AnimaIka v4.9, Hikari v1.2.0, Base (baseline) |
|---|---|---|---|
| Steps | 8 | 16 | 30 |
| CFG | 1.0 | 1.0 | 4.5 |
| Sampler / scheduler | er_sde / simple | Same | Same |
| Resolution | 832×1216 (832×1024 for bust shots, 1216×832 for two characters) | Same | Same |
| Seed | 42 | Same | Same |
| Character LoRA | keikana-animabase-v2_epoch140 (strength_model 1.0, strength_clip 1.0) | Same | Same |
| Positive prefix | masterpiece, best quality, safe | Same | Same |
“Turbo LoRA baked in” means the speed-up LoRA is already merged into the checkpoint file, and “distilled” means the model was trained to produce images in fewer steps.
For the two Turbo-baked models and the distilled one, I lowered steps and CFG to match the authors’ recommendations. The rest use the same 30 steps as the baseline. Some models recommend Euler a, but I used er_sde for all of them so they stay comparable.
I did not add @style to RDBT, to keep its prompts the same as the other models.
| Environment | Details |
|---|---|
| Machine | M1 Max 64GB |
| ComfyUI | v0.30.1, my usual install |
Default Style
No character LoRA here. I used a prompt for a blonde girl in a white robe.
Test 1 (Standing on a White Background)
| Model | What I saw in the image |
|---|---|
| Base | Strongly yellow gold, with a front panel and hem. Blunt bangs, flat shading |
| Hotaru Blend v6.0 | Large eyes and a round face. The hem flares far behind her, high saturation |
| CapAnima Turbo 1.2 | Thick lines, flat colors, chibi proportions. Emphasis lines beside the face |
| RDBT v2.3 16s | Semi-realistic with painterly shading and a more mature face. The robe looks grayish, with fine gold embroidery |
| AnimaYume v1.5 | Gold embroidery over the whole robe, the most gold of any model |
| Anima Cat Tower v1.1 | Blunt bangs. A gold bow at the chest, tassels on the sash, white boots |
| AnimaIka v4.9 | Hair pulled back from the forehead. A wide gold front panel and thick gold bands on the sleeves and hem, lots of gold |
| Hikari Anima v1.2.0 | Hair flowing to the side, fine gold embroidery on the capelet and hem. Glossy shading |
Only RDBT came out semi-realistic with painterly shading instead of anime-style cel shading. I did not include @style, so this is probably the “styles mix at random” behavior the author describes.
CapAnima produced chibi proportions even without a character LoRA. The author writes that LoRAs were merged in for a chibi style.
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 (Running Through a Meadow)
| Model | What I saw in the image |
|---|---|
| Base | The cape lining turns red. The castle is small and far away |
| Hotaru Blend v6.0 | A large castle on the left, cumulonimbus clouds on the right, white-tipped grass in front. The pose looks more like walking than running |
| CapAnima Turbo 1.2 | Still chibi, with a glaring expression. Castle on the left, sunset clouds |
| RDBT v2.3 16s | Semi-realistic, like an oil painting. A castle on a hill and dark clouds, a navy cape over the robe |
| AnimaYume v1.5 | Leaning semi-realistic. A small castle in the back right, the gold sash blowing in the wind |
| Anima Cat Tower v1.1 | Stays in anime style. A castle with a blue roof in the back left, pink clouds, white flowers in front |
| AnimaIka v4.9 | Stays in anime style. A white castle in the back right, pink and navy clouds. A blue gem on the sash knot |
| Hikari Anima v1.2.0 | Nearly turned away, in profile, arm raised. Castle in the back right, red clouds. The embroidery looks closer to silver |
Besides RDBT, AnimaYume also leaned semi-realistic in Test 2, a different look from its Test 1 result.
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 fieldSame negative as Test 1.
Character LoRA
Here I applied the character LoRA trained on Anima-Base. It is my Kei-Kana dual-character LoRA v2, made for two original characters, and kanachan / keichan in the prompts are its trigger words.
Kana-chan’s features are brown hair, an ahoge, a side ponytail on the viewer’s right, a blue scrunchie, brown eyes, and a white shirt with a red tie.
Kei-chan has long blonde hair, blunt bangs, side braids, a blue ribbon in the back of her hair, blue eyes, and a red ribbon tie.
Kana-chan
Full-body standing shots.
For the bust shots, I changed the prompt to upper body, portrait.
| Model | What I saw in the image |
|---|---|
| Base | All features present. Black tights in the full-body shot, navy blazer with a large red ribbon in the bust shot |
| Hotaru Blend v6.0 | All features present. White knee socks in the full-body shot |
| CapAnima Turbo 1.2 | Hair and eye features present. A red bow at the neck in both shots. Leaning forward with hands on her knees |
| RDBT v2.3 16s | All features present. Black tights in the full-body shot, navy blazer in the bust shot |
| AnimaYume v1.5 | All features present. Navy blazer with a thin red ribbon in the bust shot |
| Anima Cat Tower v1.1 | All features present. Navy blazer with a red bow in the bust shot |
| AnimaIka v4.9 | Features present. A thin red ribbon at the neck in the full-body shot, a red bow in the bust shot |
| Hikari Anima v1.2.0 | Features present. A thin red ribbon at the neck in both shots. Brown plaid skirt in the full-body shot |
All seven models rendered Kana’s hairstyle as the right character.
The AnimaYume v1.5 model page warns that LoRAs trained on Anima v1.0 may break, but in this one image the brown hair, ahoge, side ponytail, and blue scrunchie all came through.
Kei-chan
Full-body standing shots.
For the bust shots, I changed the prompt to upper body, portrait.
| Model | What I saw in the image |
|---|---|
| Base | Hair features and the red ribbon tie both present. Plaid skirt |
| Hotaru Blend v6.0 | Hair present, but the ribbon at the neck turns navy |
| CapAnima Turbo 1.2 | Hair features present. The ribbon is red in the full-body shot and navy in the bust shot |
| RDBT v2.3 16s | Hair features and the red ribbon tie both present. Plaid skirt |
| AnimaYume v1.5 | Hair present, but the ribbon at the neck turns blue |
| Anima Cat Tower v1.1 | Hair features present. The ribbon is red in the full-body shot and navy in the bust shot |
| AnimaIka v4.9 | Hair present, but the ribbon at the neck turns navy |
| Hikari Anima v1.2.0 | Hair present, but the ribbon at the neck turns navy. Navy jumper skirt in the full-body shot |
All seven models also got Kei’s hairstyle right.
Among the derivatives, only RDBT kept the ribbon at her neck red in both the full-body and bust shots. CapAnima and Cat Tower v1.1 were split: red in the full-body shot, navy in the bust shot.
Hotaru, AnimaYume, AnimaIka, and Hikari turned it navy or blue in both shots, the same as the official Turbo v1.0 and v1.1 in the October 21 post. The blue ribbon in the prompt (meant for the ribbon in her hair, not the one at her neck) may be pulling the color, but with one image per model at seed 42, I can’t tell that apart from random variation.
Prompts (Kana and 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 the bust shots, I replaced full body, standing with upper body, portrait and added full body, knees, feet, lower body to the negative.
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 Characters
No touching poses. I placed them next to each other with standing side by side, full body.
The prompt also gives each character’s position (left or right) and appearance in English sentences.
| Model | What I saw in the image |
|---|---|
| Base | Kei on the left, Kana on the right, no ahoge bleeding between them. Kei has a red ribbon, Kana wears black tights |
| Hotaru Blend v6.0 | Same separation and ahoge as Base. Framed from the waist up instead of full body. Kei has a navy ribbon |
| CapAnima Turbo 1.2 | Same separation and ahoge as Base. Both are chibi. Kei has a thin navy ribbon, Kana smiles with a snaggletooth |
| RDBT v2.3 16s | Grew to four people: two Keis and two Kanas. Each one is unmixed. Both Keis have red ribbons |
| AnimaYume v1.5 | Same separation and ahoge as Base. Kei has a blue ribbon, Kana wears short sleeves and black tights |
| Anima Cat Tower v1.1 | Same separation and ahoge as Base. Kei has a red bow, Kana has a red tie and black tights |
| AnimaIka v4.9 | Same separation and ahoge as Base. Kei has a thin red ribbon, Kana has a red tie and black tights |
| Hikari Anima v1.2.0 | Same separation and ahoge as Base. Kei has a thin red ribbon and brown plaid, Kana has a red tie and black tights |
In RDBT, Kei’s ribbon stayed red, but the two-character prompt doubled the number of people.
In the October 21 post, the official Aesthetic v1.1 also added extra people in the same way. Base did not.
Hotaru cut the image off at the waist even with full body in the prompt.
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: the same as for Kana and Kei, plus 1girl, solo.
Background Density
Kana-chan sits by the window in a café.
| Model | What I saw in the image |
|---|---|
| Base | Front-facing, holding a cup in both hands. Two ceiling lamps, curtains, chairs in the back |
| Hotaru Blend v6.0 | Both hands at her chin. Window, three potted plants, a bookshelf, ceiling lamps, picture frames, chairs in the back |
| CapAnima Turbo 1.2 | Holding a cup to her mouth, with a second cup on the table. Window, curtains, two potted plants, a bookshelf |
| RDBT v2.3 16s | Chin in hand, smiling. A teapot and cups on a shelf, tables and chairs in the back. Navy cardigan with a red ribbon |
| AnimaYume v1.5 | Both hands at her chin. Curtains, two ceiling lamps, picture frames, tables and chairs in the back |
| Anima Cat Tower v1.1 | Both hands on her cheeks. Two bookshelves, a wall lamp, three picture frames, two potted plants. Navy blazer with a red tie |
| AnimaIka v4.9 | Holding a cup to her mouth, with a second cup on the table. Curtains, two ceiling lamps, chairs in the back. A white border appeared around the image |
| Hikari Anima v1.2.0 | Holding a cup to her mouth, with a second cup on the table. Four potted plants, a hanging planter, a bookshelf |
All seven drew bookshelves and potted plants, with about as much background detail as Base.
CapAnima, AnimaIka, and Hikari drew a second cup on the table besides the one in her hand, so the single word teacup produced two cups.
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)
As in the April WAI-Anima post, I took a source image made with WAI-Illustrious v160 and the Kana-chan LoRA, and ran it through image-to-image (i2i) at denoise 0.5.
No character LoRA was applied.
| Model | What I saw in the image |
|---|---|
| Base | Composition, circle, and sash knot unchanged, with a pattern added inside the circle |
| Hotaru Blend v6.0 | Composition kept, a floral pattern in the circle. A white hood is added |
| CapAnima Turbo 1.2 | Composition kept, lines and colors change to flat chibi-style shading. The side ponytail becomes curled |
| RDBT v2.3 16s | Composition kept, a green gem appears at the chest. Subtle pattern in the circle |
| AnimaYume v1.5 | Composition kept, the inside of the circle becomes a text-like pattern |
| Anima Cat Tower v1.1 | Composition kept, a white hood is added. The background turns pale pink |
| AnimaIka v4.9 | Composition kept, a white hood is added. The scrunchie turns green, and a green gem appears at the chest |
| Hikari Anima v1.2.0 | Composition kept, a white hood is added. The background turns pale pink |
All eight models, Base included, kept the source image’s composition, gold circle, and sash knot. What changed was the pattern inside the circle, the white hood and chest gem, and CapAnima’s lines and shading.
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 bodySame negative as the default style tests.
Generation Speed
Seconds from when the generation script submits a job until it receives the image, on the same M1 Max 64GB.
Test 1 is the first image for each model and includes model loading. Base values come from the October 21 post.
| Model | Steps | Test 1 | Test 2 | Bust shot (832×1024) |
|---|---|---|---|---|
| Base | 30 | 262.8s | 258.7s | 210.7s |
| Hotaru Blend v6.0 | 8 | 40.2s | 38.1s | 30.2s |
| CapAnima Turbo 1.2 | 8 | 40.2s | 38.2s | 32.1s |
| RDBT v2.3 16s | 16 | 72.6s | 72.3s | 58.3s |
| AnimaYume v1.5 | 30 | 261.2s | 259.2s | 207.0s |
| Anima Cat Tower v1.1 | 30 | 263.0s | 263.0s | 210.8s |
| AnimaIka v4.9 | 30 | 263.1s | 261.0s | 208.8s |
| Hikari Anima v1.2.0 | 30 | 263.0s | 261.0s | 208.9s |
Hotaru and CapAnima run at 8 steps and CFG 1, so they matched Turbo v1.1 from the October 21 post (38.2s in Test 2). RDBT took about twice as long because of its 16 steps.
The four 30-step models took about the same time as Base.
Which One to Use
Assuming you apply a character LoRA trained on Base, this is how the results sort by use case.
| Use case | Best fit | Why |
|---|---|---|
| Quick output with a character LoRA | Hotaru Blend v6.0 | About 38 seconds per image, and both Kana’s and Kei’s hairstyles come out. Kei’s ribbon turns navy |
| Chibi art | CapAnima Turbo 1.2 | 8 steps, and the hairstyles come out in chibi proportions even with a character LoRA |
| Matching trained outfit colors too | RDBT v2.3 16s | Kei’s ribbon tie stayed red in the full-body, bust, and two-character shots. But two characters grew to four |
| Two characters side by side | The five models other than RDBT and Hotaru | Stayed at two people, one left and one right. Hotaru didn’t render full body, and RDBT added people |
| Keeping anime-style shading without a character | Hotaru, Cat Tower, AnimaIka | Stayed anime-style in both Test 1 and Test 2. RDBT leans semi-realistic without @style, and AnimaYume leaned semi-realistic in Test 2 |
References
- Hotaru Blend Anima (Civitai)
- CapAnima (Civitai)
- RDBT | Anima (Civitai)
- AnimaYume (Civitai)
- Anima Cat Tower (Civitai)
- AnimaIka (Civitai)
- Hikari Anima (Civitai)
- Anima (Hugging Face)
- October 2026 list of current Anima derivative checkpoints
- Testing Official Anima Turbo v1.1 and Aesthetic v1.1 on M1 Max to See If Base-Trained Character LoRAs Work
- Comparing four 2.9B Anima derivatives
Bonus NSFW Output Test
The test images below contain sensitive content, so the images themselves have been blurred.
No character LoRA. I generated under two conditions: without the nsfw tag, and with it (plus straight-on for a front-facing composition).
I removed safe from the positive prefix.
AnimaYume v1.5 and CapAnima Turbo 1.2 produced bodies that looked childlike in both images, so I have not posted those, even blurred.
Without the nsfw Tag
With the nsfw Tag (straight-on)
All six models shown here, Base included, produced nudity with or without the tag.
RDBT came out in a near-photographic realistic style both times, not anime-style.
In Hikari’s tagged image, her chest was hidden by her hair and a light-flare censor covered the crotch.
Prompt
Without the 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 the tag
masterpiece, best quality, 1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white backgroundSame negative as the default style tests.