<h1 id="qwen-3.5-4b-text-encoder-for-anima-2b">Qwen 3.5 4B Text Encoder for Anima 2B</h1>NEW → Now supported on Forge Neo (sd-webui-forge-neo) as a native extension! See the Forge Neo install instructions below.<h2 id="installation">Installation</h2><h3 id="comfyui">ComfyUI</h3>Clone the repo into your ComfyUI <code>custom_nodes</code> folder:<pre><code>cd ComfyUI/custom_nodes
git clone https://github.com/GumGum10/comfyui-qwen35-anima.git
</code></pre>Then restart ComfyUI.<h3 id="forge-neo">Forge Neo</h3>Clone or copy the extension into your Forge Neo <code>extensions</code> folder:<pre><code>cd sd-webui-forge-neo/extensions
git clone https://github.com/GumGum10/sd-forge-qwen-35-encoder.git
</code></pre>Then restart Forge Neo. Dependencies (<code>transformers</code>, <code>safetensors</code>) install automatically on first launch.<hr /><h2 id="what-is-this">What Is This?</h2>A drop-in upgrade for Anima 2B's text encoder. The stock Anima ships with a tiny 0.6B parameter text encoder — it works, but it struggles with complex prompts. This replaces it with a 4B parameter encoder that understands your prompts significantly better.The trade-off: the larger encoder needs alignment work to "speak the same language" as the diffusion model. We've done that work and ship the alignment files with this release. You just need to place files in the right folders and toggle a couple of settings.<hr /><h2 id="what-you-get">What You Get</h2>Pros:<ul><li>Much better understanding of complex/long prompts (7× more parameters dedicated to reading your text)</li><li>Better handling of detailed scene descriptions, multiple subjects, and nuanced instructions</li><li>Alignment controls let you blend between raw 4B output and 0.6B-compatible output</li></ul>Cons:<ul><li>Uses more VRAM than the stock 0.6B encoder (~4GB vs ~0.6GB for the text encoder portion)</li><li>Slightly slower encoding (more parameters to run)</li><li>Alignment is an approximation — the diffusion model was trained against the 0.6B, so we're rotating the 4B's output to match. It's very good (0.96 cosine similarity) but not identical</li><li>This is a reverse-engineered implementation — the original author's private code may differ in subtle ways</li></ul><hr /><h2 id="file-placement">File Placement</h2>All files are available at: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b">lylogummy/anima2b-qwen-3.5-4b</a><h3 id="comfyui">ComfyUI</h3>You'll download 4 files:<pre><code>ComfyUI/
├── models/
│ └── text_encoders/
│ └── qwen35_4b.safetensors ← THE TEXT ENCODER WEIGHTS
│
└── custom_nodes/
 └── comfyui-qwen35-anima/ ← THIS CUSTOM NODE FOLDER
 ├── __init__.py ← (comes with the node)
 ├── calibration_params.safetensors ← MAGNITUDE CALIBRATION
 ├── rotation_matrix.safetensors ← ALIGNMENT ROTATION
 └── qwen35_tokenizer/ ← TOKENIZER FILES
 ├── tokenizer.json
 ├── vocab.json
 └── merges.txt
</code></pre><h3 id="forge-neo">Forge Neo</h3>You only need to download 1 file — the calibration files, alignment matrix, and tokenizer are already bundled with the extension:<pre><code>sd-webui-forge-neo/
├── models/
│ └── text_encoder/
│ ├── qwen_3_06b_base.safetensors ← STOCK 0.6B (you already have this)
│ └── qwen35_4b.safetensors ← DOWNLOAD THIS
│
└── extensions/
 └── sd_forge_qwen35_encoder/ ← THIS EXTENSION
 ├── scripts/ ← (comes with extension)
 ├── lib_qwen35/ ← (comes with extension)
 ├── calibration_params.safetensors ← (bundled)
 ├── rotation_matrix.safetensors ← (bundled)
 └── qwen35_tokenizer/ ← (bundled)
</code></pre>Forge Neo note: Keep <code>qwen_3_06b_base.safetensors</code> selected in the top VAE/Text Encoder dropdown — its LLM adapter is still required. Do not put <code>qwen35_4b.safetensors</code> in that top dropdown.<h3 id="where-to-download-each-file:">Where to download each file:</h3>qwen35_4b.safetensors (both ComfyUI and Forge Neo) → Download from: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b/tree/main/text_encoders">text_encoders/</a> → Place in: <code>ComfyUI/models/text_encoders/</code> or <code>sd-webui-forge-neo/models/text_encoder/</code> → What it does: The actual 4B text encoder model weightscalibration_params.safetensors + rotation_matrix.safetensors (ComfyUI only — bundled in Forge Neo) → Download from: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b/tree/main/calibration">calibration/</a> → Place in: <code>ComfyUI/custom_nodes/comfyui-qwen35-anima/</code> → What they do: Calibration scales the 4B output to match the 0.6B's magnitude per dimension. The rotation matrix rotates the 4B's concept directions to match what the adapter expects.qwen35_tokenizer/ folder (ComfyUI only — bundled in Forge Neo) → Download from: <a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/anima2b-qwen-3.5-4b/tree/main/tokenizer">tokenizer/</a> → Place in: <code>ComfyUI/custom_nodes/comfyui-qwen35-anima/qwen35_tokenizer/</code> → What it does: The correct tokenizer (vocab=248K, NOT the default Qwen3 tokenizer) → Note: This will auto-download from HuggingFace on first use if you don't place it manually.<hr /><h2 id="how-to-use">How to Use</h2><h3 id="comfyui">ComfyUI</h3><ol><li>Add the "Load Qwen3.5 CLIP (Anima)" node (found under <code>loaders → Anima</code>)</li><li>Select <code>qwen35_4b.safetensors</code> from the dropdown</li><li>Connect the CLIP output to a CLIPTextEncode node</li><li>Use with your Anima 2B checkpoint as normal</li></ol><h3 id="forge-neo">Forge Neo</h3><ol><li>Load an Anima 2B checkpoint</li><li>Make sure <code>qwen_3_06b_base.safetensors</code> is in the top VAE/Text Encoder dropdown</li><li>In the generation tab, expand "Qwen3.5 Text Encoder (Anima)" and enable it</li><li>Select <code>qwen35_4b.safetensors</code> in the extension's Model File dropdown</li><li>Generate as normal — the extension intercepts text encoding automatically</li></ol><h3 id="recommended-settings-to-start-(both)">Recommended Settings to Start (both)</h3><pre><code>use_alignment: ON
alignment_strength: 0.5
use_calibration: OFF
output_scale: 1.0
</code></pre>That's it. Generate some images and compare against the stock 0.6B.<hr /><h2 id="tuning-guide">Tuning Guide</h2><h3 id="what-the-settings-actually-do-(plain-english):">What the settings actually do (plain English):</h3>use_alignment — Rotates the 4B's internal "compass" so that when it says "from the side" or "looking up", it points in the same direction the diffusion model expects. Without this, the 4B understands your prompt fine — it just communicates it in a way the diffusion model misreads.alignment_strength (0.0 – 1.0) — The rotation (direction fix) is always on when alignment is enabled. This slider controls how much the magnitude shifts to match the 0.6B:<ul><li>0.0 = Directions fixed, but keep the 4B's own signal strength</li><li>0.5 = Halfway blend ← start here</li><li>1.0 = Fully match the 0.6B's signal strength</li></ul>use_calibration — A finer-grained magnitude adjustment (per dimension instead of uniform). Can help, can also over-correct. Try it on and off and compare.output_scale — A simple multiplier on the final output. Leave at 1.0 unless you know what you're doing.<h3 id="recommended-workflow:">Recommended workflow:</h3><ol><li>Generate with alignment OFF first — see what the raw 4B gives you. The text understanding will be better, but poses/viewpoints may be off.</li><li>Turn alignment ON, set strength to 0.5 — generate the same prompts again. You should see better pose/viewpoint adherence while keeping the 4B's improved understanding.</li><li>Adjust strength — bump it up if spatial stuff is still off, pull it back if quality degrades.</li><li>Optionally enable calibration — compare on/off, keep whichever looks better for your use case.</li></ol><hr /><h2 id="faq">FAQ</h2>Q: Do I need both calibration AND alignment files? A: The alignment file (<code>rotation_matrix.safetensors</code>) is the most important one. Calibration is optional and supplementary. You can use alignment without calibration.Q: Will this work with any Anima 2B checkpoint? A: Yes — any checkpoint built on Anima 2B that uses the standard text encoder pipeline.Q: Does this need extra Python packages? A: For ComfyUI — no, everything ships with ComfyUI already. For Forge Neo — <code>transformers</code> and <code>safetensors</code> install automatically on first launch.Q: How much extra VRAM does this use? A: The 4B encoder weights are FP8 quantized, so roughly ~4GB for the text encoder. The stock 0.6B is under 1GB. Your total VRAM usage depends on your diffusion model + VAE + this.Q: Why not just scale the output by 10× instead of all this alignment stuff? A: Uniform scaling fixes the magnitude but not the directions. The 4B encodes "from the side" as a vector pointing in a completely different direction than the 0.6B. The rotation matrix fixes that. Scaling alone would be like shouting the wrong directions louder.Q: Is this better than the stock 0.6B? A: For text understanding — yes, meaningfully. For raw out-of-the-box image quality — it depends on your alignment settings and prompts. The 0.6B has the advantage of being exactly what the model was trained against. The 4B has the advantage of actually understanding complex prompts. With alignment at 0.5, most users see comparable or better results, especially on detailed prompts where the 0.6B falls short.Q: Can I use this with img2img? A: Yes — works for both txt2img and img2img on both ComfyUI and Forge Neo.Q: Why does Forge Neo still need the 0.6B model loaded? A: The Anima pipeline uses a small LLM adapter that lives on the 0.6B model. This adapter converts text embeddings into the format the diffusion model expects. The 4B provides the text understanding, but the adapter (on the 0.6B) still handles the final conversion. Both models are needed.<hr /><h2 id="credits">Credits</h2><ul><li>Anima 2B: <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima">circlestone-labs</a></li><li>Qwen 3.5 4B for Anima: <a target="_blank" rel="ugc" href="https://huggingface.co/nightknocker/cosmos-qwen3.5">nightknocker/cosmos-qwen3.5</a></li><li>Custom Node, Alignment &amp; Forge Neo Port: <a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-qwen35-anima">GumGum10</a></li></ul>

Anima 2B - Qwen 3.5 4B Text Encoder

anima2BQwen354BText_base.safetensors

BASE

<h1 id="anima-3.8b"><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/Anima-3.8B">ANIMA 3.8B</a></h1><h2 id="pro52-qwen3.5-edition">Pro52 / Qwen3.5 Edition</h2>Anima visual style powered by Qwen3.5 4B's language understanding.<code>52-block DiT</code> · <code>Qwen3.5 4B</code><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/Anima-3.8B/blob/main/assets/anima-38b.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/anima-38b.png" alt="Multi-character comparison" /></a><h2 id="how-to-use">How to use?</h2>** UPDATE **<ul><li>Now in forge neo:</li></ul><a target="_blank" rel="ugc" href="https://github.com/GumGum10/forge-anima-3.8B">https://github.com/GumGum10/forge-anima-3.8B</a><ol><li>Clone <a target="_blank" rel="ugc" href="https://github.com/GumGum10/comfyui-anima-3-8B">comfyui-anima-3-8b</a> into Comfyui -&gt; Custom nodes</li><li>Download the models from this repo and move them into difussion_models and text_encoders respectively</li><li>Restart comfy</li></ol><h2 id="nodes">Nodes</h2><h3 id="load-qwen3.5-4b-(anima)">Load Qwen3.5 4B (Anima)</h3><ul><li>select qwen35_4b.safetensors</li></ul><h3 id="qwen3.5-unified-prompt-(anima)">Qwen3.5 Unified Prompt (Anima)</h3><ul><li>plug model, native clip (0.6 - original anima clip), and qwen_35 clip into the node's INPUT</li><li>provide the prompt*</li></ul>** note on prompt: This was trained with the Description separator for NL first write your normal meta/artist/quality tags then Description: e.g:<pre><code>masterpiece, best quality, high quality, newest, year 2025, year 2024, 
Description:
A fox-girl jumps over a high fence.
</code></pre><ul><li>Select the anima-3.8b adapter, you can leave strength at 1.0.</li><li>Use the expanded OUTPUT for positive conditioning.</li><li>Likewise it is recommended to do the same with negative prompt/conditioning.</li></ul><hr />Note on the native OUTPUT:<ul><li>This is the original qwen 0.6b conditioning</li><li>Since Anima 3.8B was trained at low res, it is recommended to plug native into the second ksampler/highres pass to avoid artifacts</li></ul><h2 id="recommended-starting-point">Recommended starting point</h2><ul><li>Resolution: 832x1216 px (or any other 1MP res)</li><li>Adapter strength: 1.0</li><li>CFG: 7–8</li><li>Steps: 28–50</li><li>Sampler/scheduler: res_multistep + Beta (generally anything with Beta)</li></ul>At <code>adapter_strength: 0</code>, 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.<h2 id="prompting">Prompting</h2>Natural language, tags, and hybrids all work. For complex scenes, assign each subject its own sentence:<pre><code>Miku on top right.
Teto on lower left.
Apple on top left.
Nothing on lower right.
etc.
</code></pre>Try to avoid pronouns when working with multiple characters, e.g: she/he, use explicit names instead; e.g: Instead of<pre><code>2girls, Miku is sitting near Teto, she is eating an ice-cream.
</code></pre>Try:<pre><code>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.
</code></pre>Other than this, prompt format should follow Anima/Anima 2.9 guidance.<h2 id="what-is-this">What is this?</h2>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:<ul><li>an expanded 52-block diffusion transformer;</li><li>a progressive Qwen3.5 cross-attention adapter;</li><li>the separate Qwen3.5 4B text encoder;</li></ul>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.<blockquote>“3.8B” is the release name for the expanded Pro52 diffusion model. Qwen3.5 4B remains a separate inference component.</blockquote><h2 id="showcase">Showcase</h2>01 — Prompt adherence + typography<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/1.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/1.png" alt="Prompt adherence + typography" width="100%" /></a>02 — Spatial composition + action<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/2.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/2.png" alt="Spatial composition + action" width="100%" /></a>03 — Two-character conflict + lettering<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/3.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/3.png" alt="Two-character conflict + lettering" width="100%" /></a>04 — Exact count + object binding<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/4.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/4.png" alt="Exact count + object binding" width="100%" /></a>05 — Two-character interaction<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/5.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/5.png" alt="Two-character interaction" width="100%" /></a>06 — Costume + environment + text<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/6.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/6.png" alt="Costume + environment + text" width="100%" /></a>07 — Four-panel binding<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/7.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/7.png" alt="Four-panel binding" width="100%" /></a>08 — Two-character pose + hand interaction<a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/grids/8.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/grids/8.png" alt="Two-character pose + hand interaction" width="100%" /></a><h3 id="prompts-used-in-the-grids">Prompts used in the grids</h3>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<pre><code>(@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"
</code></pre>02 — Spatial composition + action<pre><code>(@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
</code></pre>03 — Two-character conflict + lettering<pre><code>(@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
</code></pre>04 — Exact count + object binding<pre><code>(@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
</code></pre>05 — Two-character interaction<pre><code>(@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.
</code></pre>06 — Costume + environment + text<pre><code>(@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".
</code></pre>07 — Four-panel binding<pre><code>(@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.
</code></pre>08 — Two-character pose + hand interaction<pre><code>(@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.
</code></pre><h3 id="solo-outputs">Solo outputs</h3><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/1.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/1.png" alt="Solo output 1" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/2.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/2.png" alt="Solo output 2" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/3.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/3.png" alt="Solo output 3" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/4.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/4.png" alt="Solo output 4" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/5.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/5.png" alt="Solo output 5" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/6.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/6.png" alt="Solo output 6" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/7.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/7.png" alt="Solo output 7" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/8.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/8.png" alt="Solo output 8" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/9.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/9.png" alt="Solo output 9" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/10.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/10.png" alt="Solo output 10" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/11a.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/11a.png" alt="Solo output 11a" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/13.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/13.png" alt="Solo output 13" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/14.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/14.png" alt="Solo output 14" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/15.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/15.png" alt="Solo output 15" width="100%" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/lylogummy/assets/solo/16.png"><img src="https://huggingface.co/lylogummy/Anima-3.8B/resolve/main/assets/solo/16.png" alt="Solo output 16" width="100%" /></a>The grids compare native Anima with the paired model at different Qwen strengths. They are qualitative evidence, not a promise that every seed improves.<h2 id="training">Training</h2><ul><li>Trained on highly efficient booru dataset containing all tags &gt;5% occurence</li><li>25% Natural-language, 25% booru-tag, and 50% dual/hybrid caption views</li><li>40h on a single 4090</li><li>Batch size 56 x 10h [256x256 res]</li><li>Batch size 28 x 30h [512x512 res]</li></ul><h2 id="architecture">Architecture</h2><pre><code>prompt ─┬─ Qwen3 0.6B ─ native Anima adapter ────┐
 └─ Qwen3.5 4B ─ progressive cross-attn ──┴─ 512-token conditioning
 │
 Pro52: 40 native + 12 trained DiT blocks
 │
 image
</code></pre>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.<h2 id="files">Files</h2><pre><code>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
</code></pre><h2 id="limitations">Limitations</h2><ul><li>The model and adapter must be used together.</li><li>Qwen3.5 4B adds meaningful VRAM and latency.</li><li>Exact identity, counting, hands, readable lettering, and crowded layouts can still fail.</li></ul><h2 id="licenses">Licenses</h2>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.

Files with SHA256: ea289be7c916726d09953c7db9971c82b280e694b5d7c47f8ad9ffad6acb54ba