Testing TinAnima29B 4steps INT8 on M1 Max: 12 Images with No Character Prompt Produced the Same Face

On November 13, 2026, a 4-step INT8 model based on Anima 2.9B titled “TinAnima29B_v1_4steps_INT8” was published on Civitai.
Browsing through several sample images, I noticed that despite different hair colors and clothing, the facial features looked remarkably similar.
It felt reminiscent of what was once called the “masterpiece face” back in the SD 1.5 and NovelAI eras—a default face that models drifted toward when quality tags were added.

The official Anima model card highlights flexibility and diversity as key strengths of the Base model.
So, does this derivative produce the exact same face even when no character tags are specified?
Including the negative prompt instructions provided by the author, I first verified that I could reproduce the author’s sample images locally, and then generated outputs using the same test criteria as my article comparing the 2.9B version of One obsession.

Basic Information and Recommended Settings

On November 15, 2026, I checked the model page and Civitai API details.

Although Civitai lists the base model as “Illustrious”, inspecting the safetensors header reveals 40 DiT blocks (numbered 0 to 39), identical to Anima-2.9B.
The 2.9B version of One obsession also showed this same display, so I suspect it is simply due to limited base model category choices when registering Anima 2.9B on Civitai.
The accompanying text encoder and VAE have matching SHA256 hashes with the original files distributed by CircleStone Labs, confirming they are unmodified.

The author’s recommended settings gathered from the description and sample metadata:

Interestingly, while the description specifies a negative prompt, none of the author’s 88 samples include any negative prompt.
This behavior relates to how CFG 1 operates, which I verified in a later section.

Creation Workflow

ComfyUI workflow metadata was left intact inside the safetensors file:

The same workflow also contained an INT8 save node, but it was left disabled.
I assume the INT8 quantization conversion was performed separately.
What the base model “A5” and strength 1.0 LoRA “A15_29B” actually are is not disclosed on the model page or in any other releases by the author.
The fact that character and woman LoRAs bearing the author’s name were merged at strength 0.4 each is relevant to the facial consistency experiments below.

Why this model can generate complete images in just 4 steps could not be determined from the model page.
Official Anima Turbo LoRA recommends 8 to 12 steps at CFG 1, and community Turbo models for 2.9B recommend 8 steps.

Licensing

The underlying base, Anima-2.9B, is not an official release from CircleStone Labs.
It is an unofficial community derivative where Gazingstars123 expanded official Anima (2B, 28 blocks) to 40 blocks and trained the 12 added blocks. Its license is the same CircleStone Labs Non-Commercial License as official Anima.

In contrast, TinAnima’s Civitai license toggles permit everything: selling generated images, generation services on Civitai and third-party platforms, selling the model, selling merges, requiring no attribution, and redistributing under different licenses.
Comparing this against the underlying base license reveals several discrepancies:

Allowing generation on Civitai (RentCivit) matches CircleStone’s own settings on the official Anima Civitai page.
The “no attribution” toggle is also identical on the official Anima Civitai page, reflecting Civitai’s somewhat coarse permission options.
When distributing or using TinAnima, one should rely on the underlying base license text rather than the Civitai display toggles.

Running on M1 Max

Because 40-block models cannot be loaded in my regular ComfyUI environment (v0.30.1), I used the v0.33.3 environment prepared in my article verifying Anima-2.9B LoRA compatibility.
The initial test run stopped immediately during sampler execution:

NotImplementedError: The operator 'aten::_int_mm' is not currently implemented for the MPS device.

This INT8 safetensors file contains per-tensor quantization configurations: {"format":"int8_tensorwise","convrot":true,"convrot_groupsize":256}.
ComfyUI’s quantization library (comfy_kitchen) attempts to compute these linear layers using INT8 matrix multiplication (torch._int_mm).
This acceleration path is built for CUDA, and the current PyTorch MPS backend for Apple Silicon GPUs does not implement this operator.

When setting up this environment in August, I had patched around this issue.
However, when comfy_kitchen was updated to 0.2.31 in October, that patch was overwritten.
This time, I added a branch in comfy_kitchen/tensor/int8.py for linear layer processing: if the input tensor device is MPS, dequantize the weights back to bf16 and compute via standard linear layers.

# MPS patch: torch._int_mm is not implemented on MPS -> dequantize, bf16 linear
if input_tensor.device.type == "mps":
    return torch.nn.functional.linear(*dequantize_args(args), **dequantize_args(kwargs))

The dequantization routine also inverts the Hadamard rotation (convrot) applied during quantization.
Under this patch, INT8 matrix multiplication acceleration is unavailable.
On M1 Max, the only benefit is reduced file size and initial memory footprint.

Reproducing Author Samples

To avoid falsely assuming failure due to arbitrary prompt choices, I selected 3 sample images from the author and tested whether identical images were produced using the exact same prompts and seeds.
Settings: ER SDE, Simple scheduler, 4 steps, CFG 1, 1024×1536, no negative prompt.

The sample metadata specifies “Shift 1”.
ComfyUI defaults Anima’s shift to 3.0, so I generated outputs both at default shift 3.0 and with shift set to 1.0 via ModelSamplingAuraFlow.

148870975 · shift 3.0

148870975 · shift 1.0

148870978 · shift 3.0

148870978 · shift 1.0

148872261 · shift 3.0

148872261 · shift 1.0

Across Forge Neo and ComfyUI, the same seeds produced nearly identical images.
Because shift 1.0 was closer to the author’s samples across all three, subsequent generations used shift 1.0.
At 1024×1536 4-step, each image took roughly 50 seconds on M1 Max.
Compared to the author’s ~5.5s on an RTX 3060 12GB, the lack of hardware INT8 matrix multiplication on MPS creates a significant performance gap.

Negative Prompt Behavior

Using the same conditions as sample 148870975, I tested the negative prompt listed in the description.

CFG 1 · No negative

CFG 2 · No negative

CFG 2 · Author's negative

When CFG is 1, ComfyUI bypasses negative prompt conditioning passes entirely.
As long as the model is used at the recommended CFG 1, whatever is written in the negative prompt makes zero difference.
Increasing CFG to 2 forces negative evaluation passes, nearly doubling generation time.
In this test, CFG 2 only oversaturated colors without meaningful difference whether negative tags were present or not.
The complete absence of negative prompts across all of the author’s sample metadata aligns with this behavior.

worst quality, low quality, score_1, score_2, score_3, artist name

Does a Default Face Emerge Without Character Tags?

I tested seeds 1 through 4 with a minimal prompt containing no character names, hair colors, eye colors, or artist tags.
For comparison, I used stock Anima-2.9B and Anima-Base v1.0, both configured at 30 steps and CFG 4.5 matching the One obsession article.
I also compared with and without quality tags, and tested TinAnima with the author’s suggested natural language format.
Resolution was set to 832×1024; TinAnima used 4 steps, CFG 1, and shift 1.0.

TinAnima

Tags (seeds 1–4)

Natural Language (seeds 1–4)

Quality Tags (seeds 1–4, blurred for publication due to revealing clothing)

Stock Anima-2.9B

Tags (seeds 1–4)

Quality Tags (seeds 1–4, blurred for publication due to revealing clothing)

Anima-Base v1.0

Tags (seeds 1–4)

Quality Tags (seeds 1–4, blurred for publication due to revealing clothing)

Whether using tags, natural language, or quality tags, all 12 TinAnima images produced the exact same black-haired female character.
In contrast, stock Anima-2.9B and Anima-Base generated varied hair colors, styles, and faces across seeds 1–4 when unspecified.
Under these conditions, this single fixed face in TinAnima is not an inherited trait from base Anima, but appears to have been introduced during the fine-tuning/merge process.
Even in the author’s samples with pink or white hair, the eye shapes and nose structures share this exact facial geometry.

Adding quality tags to both base models resulted in revealing, bust-emphasizing outputs across all 4 seeds.
Without the safe tag, adding masterpiece, best quality frequently biases outputs toward NSFW.
TinAnima also trended toward open necklines with quality tags, though slightly less aggressively than the base models.

Specifying Hair and Eyes Only

Next, I specified only hairstyle, hair color, and eye color without naming any character, generating seeds 1–4 in TinAnima.

While hair and eye colors updated as prompted, the underlying facial structure did not change.
The TinAnimaCharactersLora_29B and TinAnimaWomanLora_29B baked into the model at strength 0.4 each represent primary candidates for this bias.
Because neither LoRA is publicly available, I cannot isolate their individual contributions locally.

1girl, solo, short hair, blonde hair, blue eyes, looking at viewer, upper body, white background

worst quality, low quality, artist name, blurry, jpeg artifacts, chromatic aberration

Comparison Targets

From here on, I generated outputs using the same test scenarios as the One obsession article:

Generation Settings

Tests were conducted on an Apple Silicon M1 Max 64GB.
For TinAnima, matching the author’s samples, I dropped only masterpiece, best quality from the previous articles’ prompts.
safe was retained because it governs image content rather than acting as a generic quality booster.
The negative prompt used the same string as prior articles, though at CFG 1 it has no effect.
Resolution was aligned to 832×1216 matching prior articles rather than the author’s recommended 1024×1536.

Character LoRA Strength

The character LoRAs used were the 40-block remapped versions of Kei-chan and Kana-chan originally trained on Anima-Base—the identical files used in the 2.9B One obsession article.
To verify whether applying LoRA at 4 steps caused distortion, I first evaluated Kana-chan’s full-body generation across different strengths.

No LoRA

Strength 1.0

Strength 0.8

Strength 0.6

Because strength 1.0 showed no signs of degradation, I proceeded with strength 1.0 matching previous articles.

Base Art Style

Testing without any character LoRA using a blonde hair + white robe prompt.

White Background Full Body

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima rendered gold ornaments more conservatively than the other three models.
Linework is delicate, and color shading is softer.

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 background

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 background

worst quality, low quality, artist name, blurry, jpeg artifacts, chromatic aberration

Running Across a Grass Field

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima rendered a softer sunset palette compared to the deep reddish skies of the other three.
The character stands facing forward rather than exhibiting a dynamic running motion.

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 field

Character LoRA Evaluation

Testing reproduction using Kana-chan (brown side ponytail) and Kei-chan (blonde long hair).

Kana-chan

Full body:

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

Bust portrait (prompt changed to upper body, portrait):

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima accurately reproduced Kana-chan’s ahoge, scrunchie side ponytail, and brown eyes.
Where the other three produced white shirts for full-body, TinAnima rendered a navy blazer.
Eyes are somewhat wider horizontally with thicker eyelashes compared to the other models.

Kei-chan

Full body:

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

Bust portrait (prompt changed to upper body, portrait):

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima captured Kei-chan’s braid, blunt bangs, and blue hair ribbon.
The neck ribbon rendered as navy in both full body and bust, as blue ribbon in the prompt bled into the collar ribbon.
Her gaze in the bust portrait appears gentler than the sharper upturned eyes in the other models.

safe, 1girl, solo, kanachan, brown hair, side ponytail, ahoge, brown eyes, school uniform, full body, standing, looking at viewer, white background, simple background

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 background

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 face

Two Characters Standing Side by Side

Specifying standing side by side, full body with left/right positions and individual character traits:

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima was the only model that framed the composition from the chest up despite explicit full body prompts.
However, character differentiation was completely successful, with no trait blending between hairstyles or eye colors.

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 background

Background Detailing

Kana-chan seated by the window in a cafe:

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima accurately rendered background elements including books on shelves and outdoor architecture.
Framing is closer than the others, showing from the tabletop upward.
Mirroring what appeared with 2.9B v2 in the One obsession article, an extra teacup appeared on the table.

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 viewer

worst quality, low quality, blurry, lowres, jpeg artifacts, bad hands, bad anatomy, deformed face

Image-to-Image (i2i)

As in my WAI-Anima article, I used a source image generated by WAI-Illustrious v160 with Kana-chan LoRA, running i2i at denoise 0.5 without any character LoRA.
Because TinAnima runs in 4 steps, denoise 0.5 executes only the final 2 denoising steps.

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

All four models preserved the overall composition, golden halo, and sash knot from the source image.
TinAnima rendered a slightly more mature gaze compared to Kana-chan’s rounder eyes in the source.

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 body

Generation Speed

Measured on an Apple Silicon M1 Max 64GB from job start to image return.
Base, Turbo v1.1, and Anima-2.9B reflect measurements from prior articles; Test 1 includes initial model load time.
Because TinAnima was evaluated directly after the LoRA strength tests, Test 1 does not include cold-load overhead.
TinAnima computed via the MPS patch dequantizing back to bf16 rather than native INT8 matrix multiplications.

TinAnima generated images roughly 12 times faster in Test 2 compared to 30-step Anima-2.9B on the same 40-block architecture.
It is slightly faster than 8-step 2B Turbo v1.1.
At the author’s recommended 1024×1536 resolution, each image took approximately 50 seconds.

Recommendations by Use Case

References

• TinAnima29B_v1_4steps_INT8 (Civitai)
• Anima (Hugging Face)
• Anima-2.9B (Hugging Face)
• NVIDIA Open Model License
• Testing One obsession Anima, 2.9B, and Illustrious with Identical Prompts and Seeds
• Can Existing Character LoRAs Run on 40-Layer Anima-2.9B via Layer Expansion?
• Testing Official Anima Turbo v1.1 and Aesthetic v1.1 on M1 Max: Can Base-Trained Character LoRAs Be Used?
• Tracking the Evolution of the Anima Ecosystem with WAI-Anima

Bonus: NSFW Output Test

Images beyond this point are blurred because outputs from this test may be sensitive. If you wish to see unblurred results, please test in your local environment.

Tested without character LoRA across two conditions: without nsfw tag and with nsfw tag (using straight-on framing). safe was removed from prompt prefixes.

Without nsfw Tag

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

With nsfw Tag (straight-on)

Anima-Base v1.0

Anima-Turbo v1.1

Anima-2.9B-preview-v1

TinAnima29B 4steps INT8

TinAnima rendered unclothed figures regardless of whether the nsfw tag was included.
Without the tag, the character sat on a rock dipping feet into water, mirroring other Anima models.
With the tag, the character stood before a white curtain.

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 viewer

1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white background