Upload 26 files

.gitattributes

@@ -33,3 +33,28 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/anima-38b.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/1.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/2.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/3.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/4.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/5.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/6.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/7.png filter=lfs diff=lfs merge=lfs -text
+assets/grids/8.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/1.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/10.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/11a.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/12.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/13.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/14.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/15.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/16.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/2.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/3.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/4.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/5.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/6.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/7.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/8.png filter=lfs diff=lfs merge=lfs -text
+assets/solo/9.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,285 @@
+<div align="center">
+
+# ANIMA 3.8B
+
+## Pro52 / Qwen3.5 Edition
+
+**Anima visual style powered by Qwen3.5 4B's language understanding.**
+
+`52-block DiT` · `Qwen3.5 4B`
+
+</div>
+
+![Multi-character comparison](assets/anima-38b.png)
+
+## 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
+
+<a href="assets/grids/1.png"><img src="assets/grids/1.png" alt="Prompt adherence + typography" width="100%"></a>
+
+#### 02 — Spatial composition + action
+
+<a href="assets/grids/2.png"><img src="assets/grids/2.png" alt="Spatial composition + action" width="100%"></a>
+
+#### 03 — Two-character conflict + lettering
+
+<a href="assets/grids/3.png"><img src="assets/grids/3.png" alt="Two-character conflict + lettering" width="100%"></a>
+
+#### 04 — Exact count + object binding
+
+<a href="assets/grids/4.png"><img src="assets/grids/4.png" alt="Exact count + object binding" width="100%"></a>
+
+#### 05 — Two-character interaction
+
+<a href="assets/grids/5.png"><img src="assets/grids/5.png" alt="Two-character interaction" width="100%"></a>
+
+#### 06 — Costume + environment + text
+
+<a href="assets/grids/6.png"><img src="assets/grids/6.png" alt="Costume + environment + text" width="100%"></a>
+
+#### 07 — Four-panel binding
+
+<a href="assets/grids/7.png"><img src="assets/grids/7.png" alt="Four-panel binding" width="100%"></a>
+
+#### 08 — Two-character pose + hand interaction
+
+<a href="assets/grids/8.png"><img src="assets/grids/8.png" alt="Two-character pose + hand interaction" width="100%"></a>
+
+### 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.
+
+<details>
+<summary><strong>01 — Prompt adherence + typography</strong></summary>
+
+```text
+(@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"
+```
+
+</details>
+
+<details>
+<summary><strong>02 — Spatial composition + action</strong></summary>
+
+```text
+(@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
+```
+
+</details>
+
+<details>
+<summary><strong>03 — Two-character conflict + lettering</strong></summary>
+
+```text
+(@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
+```
+
+</details>
+
+<details>
+<summary><strong>04 — Exact count + object binding</strong></summary>
+
+```text
+(@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
+```
+
+</details>
+
+<details>
+<summary><strong>05 — Two-character interaction</strong></summary>
+
+```text
+(@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.
+```
+
+</details>
+
+<details>
+<summary><strong>06 — Costume + environment + text</strong></summary>
+
+```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".
+```
+
+</details>
+
+<details>
+<summary><strong>07 — Four-panel binding</strong></summary>
+
+```text
+(@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.
+```
+
+</details>
+
+<details>
+<summary><strong>08 — Two-character pose + hand interaction</strong></summary>
+
+```text
+(@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.
+```
+
+</details>
+
+### Solo outputs
+
+<p align="center">
+  <a href="assets/solo/1.png"><img src="assets/solo/1.png" alt="Solo output 1" width="32%"></a>
+  <a href="assets/solo/2.png"><img src="assets/solo/2.png" alt="Solo output 2" width="32%"></a>
+  <a href="assets/solo/3.png"><img src="assets/solo/3.png" alt="Solo output 3" width="32%"></a>
+</p>
+
+<p align="center">
+  <a href="assets/solo/4.png"><img src="assets/solo/4.png" alt="Solo output 4" width="24%"></a>
+  <a href="assets/solo/5.png"><img src="assets/solo/5.png" alt="Solo output 5" width="24%"></a>
+  <a href="assets/solo/6.png"><img src="assets/solo/6.png" alt="Solo output 6" width="24%"></a>
+</p>
+
+<p align="center">
+  <a href="assets/solo/7.png"><img src="assets/solo/7.png" alt="Solo output 7" width="24%"></a>
+  <a href="assets/solo/8.png"><img src="assets/solo/8.png" alt="Solo output 8" width="24%"></a>
+  <a href="assets/solo/9.png"><img src="assets/solo/9.png" alt="Solo output 9" width="24%"></a>
+</p>
+
+<p align="center">
+  <a href="assets/solo/10.png"><img src="assets/solo/10.png" alt="Solo output 10" width="48%"></a>
+  <a href="assets/solo/11a.png"><img src="assets/solo/11a.png" alt="Solo output 11" width="48%"></a>
+</p>
+
+
+
+<p align="center">
+  <a href="assets/solo/13.png"><img src="assets/solo/13.png" alt="Solo output 13" width="64%"></a>
+</p>
+
+<p align="center">
+  <a href="assets/solo/14.png"><img src="assets/solo/14.png" alt="Solo output 14" width="32%"></a>
+  <a href="assets/solo/15.png"><img src="assets/solo/15.png" alt="Solo output 15" width="32%"></a>
+</p>
+
+<p align="center">
+  <a href="assets/solo/16.png"><img src="assets/solo/16.png" alt="Solo output 16" width="64%"></a>
+</p>
+
+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
+
+```text
+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
+
+```text
+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
+```
+
+## Recommended starting point
+
+- Resolution: **832x1216 px**
+- Adapter strength: **1.0**
+- CFG: **7–8**
+- Steps: **28–40**
+- Merge mode: **adapter**
+- Sampler/scheduler: **res_multistep + Beta** (generally anything with Beta)
+
+At `adapter_strength: 0`, conditioning is exactly native. Increase it while holding the seed fixed. High CFG often makes the learned semantic branch more visible; if the output becomes harsh or characters homogenize, reduce CFG before reducing strength.
+
+## Prompting
+
+Natural language, tags, and hybrids all work. For complex scenes, assign each subject its own sentence:
+
+```text
+
+```
+
+## 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.
+
+---
+
+<div align="center">
+
+### ANIMA 3.8B
+
+**More room for the prompt to matter.**
+
+</div>