v1.1 / Semantic Connector v2

Anima visual style with Qwen3.5 4B language understanding.

52-block DiT · Qwen3.5 4B

What is this?

Anima 3.8B is an expansion of Anima 2.9B aimed at prompt adherence, multi-character binding, interactions, spatial instructions, and mixed natural-language/tag prompting. Anima 2.9B's layers were expanded to 52, making the diffusion model approximately 3.8B parameters, hence the name.

HF: https://huggingface.co/lylogummy/Anima-3.8B

Comfy Workflow: https://huggingface.co/lylogummy/Anima-3.8B/blob/main/workflows/workflow_1.1.json

Version 1.1 contains:

• the expanded 52-block diffusion transformer;

• the new timestep-aware Semantic Connector v2, bundled into the diffusion checkpoin;

• the separate Qwen3.5 4B and native Qwen3 0.6B text encoders.

“3.8B” is the release name for the expanded Pro52 diffusion model. Qwen3.5 4B remains a separate inference component.

What changed in v1.1

Trained for ~140h on 4xA40s.

Semantic Connector v2 replaces the older progressive-cross experiment. It is timestep-aware, so it remains part of the sampling model and is evaluated at each denoising step. The connector and DiT now ship as one inference checkpoint. There is no separate adapter file to select and no strength slider to tune: v2 always uses the trained strength of 1.0.

The text encoders are still separate. They are only needed when a prompt is encoded and can be released or offloaded afterward. Changing the prompt loads them again; reusing cached conditioning does not. This keeps the steady sampling footprint much lower than holding the diffusion model and both text encoders in VRAM at the same time.

The original preview material is preserved below for reference. The old v1 adapter remains supported by the companion extensions as a fallback, but v1.1/v2 is the release path going forward.

How to run v1.1

Download the bundled Anima-3.8B-v1.1.safetensors checkpoint along with:

• qwen_3_06b_base.safetensors

• qwen35_4b.safetensors

• qwen_image_vae.safetensors

ComfyUI

Install comfyui-anima-3-8B, then place the files here:

ComfyUI/models/
├── diffusion_models/Anima-3.8B-v2.safetensors
├── text_encoders/qwen_3_06b_base.safetensors
├── text_encoders/qwen35_4b.safetensors
└── vae/qwen_image_vae.safetensors

In the workflow:

1. Load the bundle with anima.3-8B-v2.

2. Load qwen_3_06b_base.safetensors with ComfyUI's CLIPLoader and select stable_diffusion as the type.

3. Load qwen35_4b.safetensors with Load Qwen3.5 4B (Anima).

4. Connect all three to anima.3-8B-v2 Prompt, enter one prompt, and send its expanded output to the sampler.

Both text encoders are released after conditioning is produced. The bundled connector stays with the diffusion model because it is used during sampling.

Forge Neo

Install forge-anima-3.8B, place the bundle in models/Stable-diffusion, and put both text encoders in models/text_encoder. Select the v2 checkpoint, then select the native Qwen encoder and Anima VAE in Forge's VAE / Text Encoder control. The extension recognizes the bundle automatically, even when its accordion is collapsed.

On lower-VRAM systems, prompt encoding may temporarily increase VRAM use while Qwen3.5 is active. Forge and ComfyUI can offload the text encoders again before sampling. Exact memory use still depends on resolution, attention backend, precision, and the frontend's offload settings.

Recommended starting point

• Resolution: 832x1216 px (or any other 1MP res)

• CFG: 4–7

• Steps: 28–50

• Sampler/scheduler: res_multistep + Beta (generally anything with Beta)

Prompting

Natural language, tags, and hybrids all work. For complex scenes, assign each subject its own sentence:

Miku on top right.
Teto on lower left.
Apple on top left.
Nothing on lower right.
etc.

Try to avoid pronouns when working with multiple characters, e.g: she/he, use explicit names instead; e.g: Instead of

2girls, Miku is sitting near Teto, she is eating an ice-cream.

Try:

There are two girls in this illustration.
Miku from Vocaloid and Teto from Vocaloid.
Miku is sitting near Teto, Miku is eating an ice-cream.

Other than this, prompt format should follow Anima/Anima 2.9 guidance.

What is this?

Anima 3.8B is an experimental expansion of Anima 2.9B aimed at prompt adherence, multi-character binding, interactions, spatial instructions, and mixed natural-language/tag prompting. Anima 2.9B's layers were expanded to 52, making this model essentially 3.8B, hence the name.

It is a paired release:

• an expanded 52-block diffusion transformer;

• a progressive Qwen3.5 cross-attention adapter;

• the separate Qwen3.5 4B text encoder;

This is not a prompt translator or alignment model for qwen 3 0.6b. Qwen3.5 hidden states condition the denoiser through learned cross-attention, while the accompanying Pro52 checkpoint contains the trained DiT blocks that consume that signal.

“3.8B” is the release name for the expanded Pro52 diffusion model. Qwen3.5 4B remains a separate inference component.

Showcase

01 — Prompt adherence + typography

02 — Spatial composition + action

03 — Two-character conflict + lettering

04 — Exact count + object binding

05 — Two-character interaction

06 — Costume + environment + text

07 — Four-panel binding

08 — Two-character pose + hand interaction

Prompts used in the grids

These prompts were extracted directly from the PNG metadata. Open a comparison image at full size to read its panel labels.

01 — Prompt adherence + typography

(@chen bin:1.1), masterpiece, best quality, high quality, newest, year 2025, year 2024,
Description:
Foxgirl, blonde hair, braids, wearing high leg shorts and a crop top. behind her is a lake with fish, fisheye. Very detailed and intricate scenery. On her left hand she holds a sign with the text: "ANIMA 3.8B". On the wooden railing behind her there is another sign with the text: "QWEN 4B"

02 — Spatial composition + action

(@chen bin:1.1), masterpiece, best quality, high quality, newest, year 2025, year 2024,
Description:
flandre scarlet, touhou, The girl is standing in an infinite mirror rooms, around the girl there is a white snake with red eyes, the girl does an action pose, holds a sword pointed at the viewer, fisheye, foreshortening, very detailed and intricate background

03 — Two-character conflict + lettering

(@chen bin:1.1), masterpiece, best quality, high quality, newest, year 2025, year 2024,
Description:
2girls, arknights: endfield, intense confrontation, emotional fight scene, dramatic action scene, close-range conflict, dynamic composition, diagonal composition, strong tension, cinematic lighting, high contrast, impact moment, motion blur, speed lines, flying debris, hair flying, clothes fluttering, dramatic shadows, emotionally charged atmosphere, close-up battle scene, clear height of emotion, best quality, amazing quality
,four large Chinese calligraphy characters in the four corners, top-left "空", top-right "是", bottom-left "即", bottom-right "色", bold brush calligraphy, powerful ink strokes, dramatic typography, stylized kanji composition, text integrated into the scene, strong visual balance, teXt: 色
即
是
空

04 — Exact count + object binding

(@chen bin:1.1), masterpiece, best quality, high quality, newest, year 2025, year 2024, safe, detailed pupils,
Description:
Exactly three girls sit around a circular café table.
Flandre scarlet from touhou sits on the left and pours tea from a white teapot.
Kagamine rin from Vocaloid sits in the center and holds a slice of strawberry cake.
Shimakaze \(kancolle\) from kantai collection sits on the right and writes in a blue notebook.
dark background

05 — Two-character interaction

(@chen bin:1.1), masterpiece, best quality, high quality, newest, year 2025, year 2024,
Description:
David martinez from cyberpunk is holding a huge sword horizontally, looking at viewer, he is wearing a high visibility vest and baggy clothes.
Rebecca from cyberpunk is sitting on the sword that David is holding, with her back to the viewer and looking back, she is wearing a black high-tech bodysuit.
Cyberpunk street view from a low angle.

06 — Costume + environment + text

(@chen bin:1.1), masterpiece, best quality, high quality, newest, year 2025, year 2024,
Description:
Ganyu \(genshin impact\), genshin impact, in a japanese garden, the girl is wearing a pink kimono with white butterfly patterns. She can be seen walking from head to toe. Underneath her are wet stones and puddle left by a passing rain. Sakura leaves surround her, gusty, windy, fisheye. She is looking at the viewer, the background and the illustration are very intricate with lots of tiny details. Behind the girl there is an old wooden wall. There is a bold cursive japanese-style text at the top saying "ANIMA 3.8B".

07 — Four-panel binding

(@chen bin:1.1),
Description:
4 panel illustration.
top left panel is emilia from re zero.
top right panel is hatsune miku from vocaloid.
bottom left panel is burnice white from zenless zone zero.
bottom right panel is a purple and white apple.

08 — Two-character pose + hand interaction

(@chen bin:1.1),
Description:
Emilia from Re:Zero and Hatsune Miku from Vocaloid stand together in a flower garden. Emilia is on the left, offering Miku a small white flower with her right hand. Miku is on the right, accepting it with her left hand while holding a microphone behind her back with her other hand. Exactly two girls, both visible from head to toe, distinct bodies, correct character designs, clear eye contact, no duplicated characters.

Solo outputs

The grids compare native Anima with the paired model at different Qwen strengths. They are qualitative evidence, not a promise that every seed improves.

Training

• Trained on highly efficient booru dataset containing all tags >5% occurence

• 25% Natural-language, 25% booru-tag, and 50% dual/hybrid caption views

• 40h on a single 4090

• Batch size 56 x 10h [256x256 res]

• Batch size 28 x 30h [512x512 res]

Architecture

prompt ─┬─ Qwen3 0.6B ─ native Anima adapter ────┐
        └─ Qwen3.5 4B ─ progressive cross-attn ──┴─ 512-token conditioning
                                                        │
                                     Pro52: 40 native + 12 trained DiT blocks
                                                        │
                                                      image

Qwen3.5 layers 7/15/23/31 provide the semantic features. Six frozen native adapter blocks hold the original Anima alignment; six learned cross-attention insertions add the semantic residual.

Files

ComfyUI/models/
├── diffusion_models/Anima-2.9B/
│   └── Anima-3.8-preview-0.1.safetensors
├── text_encoders/
│   ├── qwen_3_06b_base.safetensors
│   ├── qwen35_4b.safetensors
│   └── Anima-3.8-preview-0.1-adapter.safetensors
└── vae/
    └── Qwen2D-Anime-dense_epoch_1.safetensors

Limitations

• The model and adapter must be used together.

• Qwen3.5 4B adds meaningful VRAM and latency.

• Exact identity, counting, hands, readable lettering, and crowded layouts can still fail.

Licenses

The companion ComfyUI code is MIT. This model is derived from and depends on upstream Anima-base and Anima 2.9B licenses. Please follow the original licenses.