Testing One obsession Across Anima, 2.9B, and Illustrious Under the Same Prompt and Seed
Contents
When searching for Anima checkpoints on Civitai, two model pages appear under the name “One obsession.”
One is One obsession_Anima, and the other is the widely downloaded One obsession.
Both are created by maxfeifei8, but their underlying base models belong to completely different lineages.
The latter started as a derivative of Illustrious (an SDXL-based anime fine-tune) and evolved through v24, with an Anima 2.9B version subsequently posted on the same page.
In contrast, the former is built directly on Anima 2B (DiT architecture).
On November 1, v5.0 of the Anima version and v2 of the 2.9B version were released simultaneously.
What differences emerge when testing models released by the same creator across three distinct base lineages (Anima 2B, Anima 2.9B, and Illustrious) under identical prompts and seeds?
Using the evaluation criteria from the comparison of four recent 2.9B derivatives, I compared them side by side in the order of v5.0, 2.9B v2, and Illustrious v24.
Basic Information and Recommended Settings
The information below was verified via Civitai’s API and model pages on November 2, 2026.
| Model | Version | Release Date (UTC) | Base | Files | Creator’s Note |
|---|---|---|---|---|---|
| One obsession_Anima | v5.0 | November 1 | Anima 2B (28 blocks) | DiT core only, fp16, 3.9 GB | Modified from Anima-Turbo |
| One obsession | Anima2.9B v2 | November 1 | Anima 2.9B (40 blocks) | DiT core only, bf16, 5.44 GB | None |
| One obsession | v24 | August 11 | Illustrious | Full checkpoint with VAE and text encoder, fp16, 6.46 GB | None |
SHA256 hashes for all three files downloaded from Civitai matched the listed values.
The creator’s recommended parameters were retrieved from page descriptions and sample image metadata.
| Parameter | v5.0 | v24 (Illustrious Page) |
|---|---|---|
| Steps | 25 in sample images | 25–35 |
| CFG | 4 in sample images | 3–6 |
| Sampler | ER SDE in sample images | Euler a, Euler |
| Resolution | 832×1216 in sample images | 832×1216, 1024×1536, etc. |
| Quality Tags | masterpiece, best quality, score_9, score_8, score_7, absurdres, newest, very aesthetic, amazing quality, highres, sensitive | masterpiece, best quality, amazing quality, very awa, absurdres, newest, very aesthetic, depth of field, highres |
The sample metadata for v5.0 used 25 steps and CFG 4, rather than Turbo’s 8-step setting.
Although described as modified from Turbo, the sample images ran standard step counts.
Therefore, this test evaluated v5.0 under both 30-step and 8-step configurations.
Architecture of 2.9B v2
On Civitai, the base model field for 2.9B v2 is listed as “Illustrious,” which led to questions in the comments.
Inspecting the file header confirms it contains 40 DiT blocks (numbered 0 to 39), matching the architecture of Anima-2.9B.
No Illustrious weights are present.
In comparison, v24 is an SDXL checkpoint packaging VAE, text encoder, and UNet into 2,515 tensors, while v5.0 contains 685 tensors across 28 blocks, matching Turbo v1.1.
The safetensors header of 2.9B v2 preserved the ComfyUI workflow used during its creation:
| Step | Node | Configuration |
|---|---|---|
| 1 | UNETLoader | One obsession_anima29B |
| 2 | UNETLoader | an animaDeltaMix29B_deltamixV60Flat |
| 3 | ModelMergeSimple | Merges steps 1 and 2 with a ratio of 0.6 |
| 4 | LoraLoaderModelOnly | Applies an anima-base-1-masterpiece-v51 at strength 1.0 |
| 5 | ModelSave | Saves the model |
While the model page lacks extensive descriptions, this workflow indicates it merges an earlier 2.9B release with another 2.9B derivative, followed by a merged LoRA.
This complete sequence was preserved in the file metadata.
Comparison Targets
| Label | Model | Role |
|---|---|---|
| Base | Anima-Base v1.0 | 30-step baseline; base model for character LoRA training. Reused outputs from the Turbo v1.1 vs Aesthetic v1.1 comparison |
| Turbo v1.1 | Anima-Turbo v1.1 | 8-step baseline. Reused outputs from the same article |
| v5.0 (8) | One obsession_Anima v5.0 | 8 steps, CFG 1.0 |
| v5.0 (30) | One obsession_Anima v5.0 | 30 steps, CFG 4.5 |
| Anima-2.9B | Anima-2.9B-preview-v1 | 2.9B baseline. Generated under identical conditions for this test |
| 2.9B v2 | One obsession Anima2.9B v2 | 30 steps, CFG 4.5 |
| WAI v17 | WAI-Illustrious-SDXL v17.0 | Illustrious baseline. Generated under identical conditions for this test |
| v24 | One obsession v24 | Latest release of the Illustrious version |
Generation Settings
| Parameter | Anima (2B) | Anima (2.9B) | Illustrious |
|---|---|---|---|
| Steps / CFG | 30 / 4.5 (8 / 1.0 for v5.0 (8) and Turbo v1.1) | 30 / 4.5 | 30 / 6.0 |
| Sampler / Scheduler | er_sde / simple | er_sde / simple | euler_ancestral / karras |
| Resolution | 832×1216 (832×1024 for portrait bust, 1216×832 for dual character) | Same as left | Same as left |
| Seed | 42 | 42 | 42 |
| Character LoRA | keikana-animabase-v2_epoch140 (strength_model 1.0, strength_clip 1.0) | Same LoRA remapped for 40 blocks | Kana-chan LoRA for WAI-Illustrious: kanachan_v17_dim16_clean (strength_model 1.0, strength_clip 0.8) |
| Text Encoder / VAE | qwen_3_06b_base / qwen_image_vae | Same as left | Built into checkpoint |
| ComfyUI | v0.30.1 | v0.33.3 (natively supports 40-block models) | v0.30.1 |
All tests were run locally on an Apple Silicon M1 Max with 64 GB of unified memory.
[!NOTE] This comparison keeps the prompt text and random seed (42) identical to observe practical output behavior across models, but it is not a strictly controlled architecture benchmark. Anima and Illustrious models utilize fundamentally different generation pipelines: Anima uses the external Qwen 0.6B text encoder and Qwen Image VAE with an
er_sdesampler, whereas Illustrious runs an SDXL UNet with integrated CLIP encoders andeuler_ancestral. Results reflect each model’s real-world behavior within its recommended ecosystem rather than checkpoint weights in isolation.
Prompts used identical Danbooru tag sequences across Anima and Illustrious models, with the positive prefix standardized to masterpiece, best quality, safe.
The creator’s recommended quality tags were omitted to maintain consistency with previous comparison benchmarks.
The step count, CFG, and sampler for Illustrious models were aligned with the WAI v17 baseline settings from the WAI-illustrious-HSWQ comparison, which fall within v24’s recommended operating envelope (25–35 steps, CFG 3–6, Euler a).
Because a Kei-chan LoRA is not available for Illustrious, the two Illustrious models were evaluated solely on solo Kana-chan outputs. Character fidelity across both Kei and Kana should be compared within the Anima lineage.
Base Art Style
Tested without character LoRAs using a blonde hair and white robe prompt.
Standing Pose on White Background
| Model | Observed Output |
|---|---|
| Base | Bright yellow gold on apron and hem. Blunt bangs with flat cel shading |
| Turbo v1.1 | Wavy hair, brownish-gold sash and embroidery. Robe takes on a grayish tone |
| v5.0 (8) | Wavy hair with blunt bangs. Gold is concentrated on cuffs and hem embroidery; robe has larger white areas. Hem flutters backward |
| v5.0 (30) | Wavy hair with exposed forehead. Wide gold apron down the center, thick gold embroidery on hem. Symmetrical composition |
| Anima-2.9B | Straight hair with blunt bangs. Bright yellow-gold apron and hem, closely resembling Base |
| 2.9B v2 | Wavy hair with exposed forehead. Patterned gold apron, ruffled cuffs, hem flutters backward |
| WAI v17 | Gray background. Angled body posture with tasseled gold sash |
| v24 | Cream-toned robe with lower saturation than others. Angled body posture with tasseled gold sash |
Even from the same model file, v5.0 showed noticeable differences between 8 and 30 steps in bangs styling and embroidery placement.
2.9B v2 diverged from stock Anima-2.9B in both hairstyle and shading, with stock Anima-2.9B leaning toward Base’s flat coloring.
The two Illustrious checkpoints shared an angled standing pose and a tasseled sash.
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 aberrationRunning Scene in a Grass Field
| Model | Observed Output |
|---|---|
| Base | Red cloak lining. Castle is small and distant |
| Turbo v1.1 | Black castle silhouette on the left. Walking forward |
| v5.0 (8) | White castle in the background on the left, large cloud on the right. Walking forward |
| v5.0 (30) | Blue-roofed castle in the background on the left, sunset on the right. Walking in black boots |
| Anima-2.9B | Red cloak lining. Small castle in the distance on the right. Artifact resembling text/signature in lower right |
| 2.9B v2 | Blue-roofed castle on the right, large cumulus clouds, white flowers in the foreground |
| WAI v17 | Large black castle on the hill. Running toward viewer |
| v24 | Large castle silhouette behind the character, purple and orange sky. Heavier, painterly shading |
Anima models rendered the castle small and far in the distance, whereas the two Illustrious models placed it large and immediately behind the subject.
Anima-2.9B produced an artifact resembling a watermark or signature in the lower right corner.
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 prompt matches the standing pose.
Character LoRA
Anima-Base-trained character LoRAs were applied to the Anima models, while a WAI-Illustrious Kana-chan LoRA was used for the Illustrious models.
Character reproducibility was evaluated on brown-haired side-ponytail Kana-chan and blonde long-haired Kei-chan.
Kana-chan
Full-body standing figures were generated first:
Portrait bust outputs were tested by changing prompt specifications to upper body, portrait:
| Model | Observed Output |
|---|---|
| Base | All key traits present. Full body features black tights; portrait bust features navy blazer and large red ribbon |
| Turbo v1.1 | All key traits present. Full body features knee-high socks |
| v5.0 (8) | Hair and eye traits reproduced. Thin red ribbon at collar in full body; red bow tie in portrait bust |
| v5.0 (30) | All key traits present. Red necktie in both full body and portrait; navy blazer in portrait bust |
| Anima-2.9B | All key traits present. Red necktie and black tights in full body; red bow tie in portrait bust |
| 2.9B v2 | Hair and eye traits reproduced. Thin red ribbon in full body; navy blazer and red bow tie in portrait bust |
| WAI v17 | Hairstyle reproduced. Full body appears on gray background with short sleeves, brown skirt, sneakers, and no necktie. Portrait bust zooms too tight, clipping the ahoge |
| v24 | Hairstyle reproduced. Full body features beige vest, red bow tie, and angled body. Portrait bust clips the ahoge similarly to WAI v17 |
All six Anima models accurately depicted Kana-chan’s characteristic hair features.
Only v5.0 (30 steps) generated the red necktie across both full-body and portrait compositions.
The two Illustrious checkpoints (WAI v17 and v24) generated nearly identical portrait framing under the same LoRA and seed.
Kei-chan
Full-body standing figures:
Portrait bust figures:
| Model | Observed Output |
|---|---|
| Base | Distinctive hair traits and red ribbon tie present. Plaid skirt |
| Turbo v1.1 | Hair reproduced, but ribbon turns dark blue. Full body jacket is also navy |
| v5.0 (8) | Hair traits present. Ribbon remains red in both full-body and portrait |
| v5.0 (30) | Hair traits present. Ribbon remains red in both full-body and portrait. Skirt is gray in full body |
| Anima-2.9B | Hair traits present. Ribbon is red in full body, dark blue in portrait. Full body wears navy double-breasted blazer |
| 2.9B v2 | Hair traits present. Ribbon is red in full body, dark blue in portrait. Full body has hands folded in front |
Kei-chan’s hairstyle was reliably rendered across all six models.
Aside from Base, only v5.0 (in both 8-step and 30-step runs) maintained the red ribbon across both compositions.
As noted in the four 2.9B models benchmark, portrait prompts often drift toward a navy ribbon influenced by blue ribbon in the tag sequence.
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 backgroundPortrait bust replaced full body, standing with upper body, portrait and appended full body, knees, feet, lower body to the negative prompt.
Negative Prompt:
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 faceKei and Kana Side by Side
Specified standing side by side, full body alongside positional and visual tags for both characters.
| Model | Observed Output |
|---|---|
| Base | Kei on left, Kana on right. No ahoge bleed. Kei has red ribbon; Kana has black tights |
| Turbo v1.1 | Clean character separation and no ahoge bleed. Kei has red ribbon; Kana has red tie and black tights |
| v5.0 (8) | Clean separation and no ahoge bleed. Kei has red ribbon and gray plaid; Kana wears a backpack, red tie, and black tights |
| v5.0 (30) | Clean separation and no ahoge bleed. Kei has red ribbon and brown plaid; Kana has red tie and black tights |
| Anima-2.9B | Clean separation and no ahoge bleed. Kei has red ribbon and gray plaid; Kana has red tie and black tights |
| 2.9B v2 | Clean separation and no ahoge bleed. Kei wears beige sweater with red ribbon; Kana has red tie and black tights |
All six models cleanly separated both characters on their respective sides without ahoge contamination across subjects.
Kei-chan’s ribbon remained red in every run.
Under v5.0 (8 steps), Kana-chan generated an unprompted backpack.
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 matches the solo tests with 1girl, solo added.
Background Details
Tested with Kana-chan seated at a cafe window table.
| Model | Observed Output |
|---|---|
| Base | Front-facing, holding cup with both hands. Two ceiling lamps, curtains, chairs in background |
| Turbo v1.1 | Resting chin on hand, cup on round table. Window, two potted plants, bookshelf |
| v5.0 (8) | Resting chin on hand, cup on round table. Beige vest. Teapot and cups on shelves, two potted plants, framed picture |
| v5.0 (30) | Bringing cup to mouth; a second cup sits on the table. White border bands appear on left and right, resembling a frame |
| Anima-2.9B | Resting chin on hand, cup on table. Beige vest. Ceiling lamp, bookshelf, counter cups |
| 2.9B v2 | Bringing cup to mouth; second cup on table. Ceiling lamp, picture frame, bookshelf, teapot |
| WAI v17 | Hands clasped beneath chin. Strong window light, brown ribbon at collar. Composition crops strictly above table |
| v24 | Hands to cheeks. Cup on table, potted plants, bookshelf, picture frame. Composition crops strictly above table |
All eight models drew bookshelves and potted plants in intricate detail.
Only v5.0 (30 steps) exhibited blank white borders on the left and right margins, a behavior that did not recur in other tests.
In v5.0 (30 steps) and 2.9B v2, a second teacup appeared on the table alongside the one held.
Both Illustrious checkpoints cropped closer to the subject.
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 WAI-Anima article, an image generated with WAI-Illustrious v160 and Kana-chan LoRA served as the base input, processed at denoise 0.5 without character LoRAs.
| Model | Observed Output |
|---|---|
| Base | Composition, circular emblem, and sash knot retained; pattern added inside the emblem |
| Turbo v1.1 | Composition retained; white hood added. Pattern inside emblem softens its outline |
| v5.0 (8) | Composition retained; white hood added. Floral pattern inside the emblem |
| v5.0 (30) | Composition retained; white hood added. Pattern inside the emblem |
| Anima-2.9B | Composition retained; white hood added |
| 2.9B v2 | Composition retained; white hood added |
| WAI v17 | Composition retained; white hood added. Hair shifts toward reddish-brown; sash shows glossy sheen |
| v24 | Composition retained; white hood added. Hair shifts toward dark brown; sash shows glossy sheen |
All eight models preserved the core composition, golden circular emblem, and sash knot from the source image.
The two Illustrious checkpoints altered hair hue and sash luster to textures distinct from the Anima family.
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 matches the standing pose.
Generation Speed
Measured elapsed time on an M1 Max 64GB from submitting the generation prompt to receiving the completed image.
Test 1 includes the initial model loading time (Base and Turbo v1.1 values are from earlier tests; 2.9B runs were measured on ComfyUI v0.33.3).
| Model | Steps | Test 1 | Test 2 | Portrait Bust (832×1024) |
|---|---|---|---|---|
| Base | 30 | 262.8 s | 258.7 s | 210.7 s |
| Turbo v1.1 | 8 | 40.2 s | 38.2 s | 30.2 s |
| v5.0 (8) | 8 | 38.2 s | 38.2 s | 30.3 s |
| v5.0 (30) | 30 | 264.8 s | 263.2 s | 208.6 s |
| Anima-2.9B | 30 | 374.2 s | 368.1 s | 293.6 s |
| 2.9B v2 | 30 | 370.2 s | 368.1 s | 293.7 s |
| WAI v17 | 30 | 104.5 s | 90.4 s | 74.3 s |
| v24 | 30 | 106.5 s | 90.4 s | 74.4 s |
At matching step counts, v5.0 performed almost identically to Base and Turbo v1.1, with 8-step runs cutting generation time in Test 2 by roughly a factor of 7.
The two 2.9B models required approximately 1.4× the time of Base, while the two Illustrious models completed generations in roughly one-third of Base’s time.
How to Choose by Use Case
| Use Case | Recommended Model | Key Observations |
|---|---|---|
| Fast generation with Base-trained character LoRA | v5.0 (8 steps) | ~38 seconds per image. Maintained red ribbon for Kei-chan across both full-body and portrait shots |
| Accurate accessories down to Kana-chan’s red tie | v5.0 (30 steps) | Produced red tie across both full-body and portrait. Exhibited white margin bands in cafe scene |
| 2.9B model with styling distinct from stock release | 2.9B v2 | Produced wavy hair and ornate apron compared to stock Anima-2.9B’s flatter cel-shading |
| Speed priority while utilizing Illustrious LoRAs | v24 | ~90 seconds at 30 steps. Shared nearly identical portrait composition with WAI v17 under the same LoRA and seed |
NSFW Output Comparison Without Character LoRA
Note: The following section evaluates uncensored outputs; preview images have been blurred for publication.
Outputs were generated without character LoRAs under two conditions: without the nsfw tag and with the nsfw tag (specifying a straight-on front-facing pose). safe was removed from the positive prompt prefix.
Without the nsfw Tag
With the nsfw Tag (straight-on)
All eight models produced nude outputs regardless of whether nsfw was present in the prompt.
In the tagless configuration, Anima models seated the subject on rocks, whereas the two Illustrious models placed her on a fallen log.
Prompt
Without nsfw 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 nsfw tag:
masterpiece, best quality, 1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white backgroundNegative matches the standing pose.
Reference Links
- One obsession_Anima (Civitai)
- One obsession (Civitai)
- Anima (Hugging Face)
- Anima-2.9B (Hugging Face)
- Applying Remapped 40-Block Character LoRAs to Four Recent Anima 2.9B Derivatives
- Testing Official Anima Turbo v1.1 and Aesthetic v1.1 on M1 Max: Can Base-Trained LoRAs Run?
- Does Layer Expansion to 40 Blocks in Anima-2.9B Retain Character LoRA Compatibility?
- WAI-illustrious-HSWQ vs WAI v17: Fast Mode in ~30s on M1 Max
- Tracking the Evolution of the Anima Ecosystem Through WAI-Anima