Anima 3.8B · Qwen 0.6B
Anima-Expanded [3.8B] をベースにマージした PPP Anima 3.8B / Qwen 0.6Bです 。
ものは試しに、と興味本位でマージを重ねてみたらPPPAnima 2.7 より精彩なイメージが生成されるので折角なので公開します。
ちなみにComfyでの動作は一切フォローしません。
使ったことがなく、多分今後も使いません。
VAE / Text Encoderは
qwen_image_vae + qwen_3_0.6_base
の組み合わせで生成できます。
サンプリングステップ数は
25～30 程度が適当かと思われます。
※私以外のAI生成サービスへの転載は許可していません。
This is PPP Anima 3.8B, created by merging models based on Anima-Expanded [3.8B].
I decided to experiment out of curiosity and tried merging them repeatedly, and since it generated more vivid images than PPPAnima 2.7, I thought I’d share it while I have the chance.
The VAE / Text Encoder can be generated using either of the following combinations:
qwen_image_vae + qwen_3_0.6_base
Generation is possible with combination.
The recommended number of sampling steps is
around 25 to 30.
*Reposting to AI-generation services other than mine is not permitted.
有馬かな、かわいいんだよ！
Description
Anima 3.8B をベースに PPP Anima 2.7なモデルです。
PPP Anima 2.7よりくっきりとした仕上がりで生成されます。
(と、自分では感じてる)
This model is based on Anima 3.8B and uses the PPP Anima 2.7 architecture.
It produces results with sharper details than PPP Anima 2.7.
(At least, that's how I see it.)
FAQ
Comments (9)
早い！ありがとうございます。ギネスビールに気づきました。。ｗ ：）
ポーターとかスタウトとか大好物です ^^
Could you please share your workflow using Qwen3.5 4B with this model?
I use only Forge Neo.
Never use Comfy, sorry...
@PPP_sleepy But you mention an Anima-3.8B-expanded_adapter which is needed when using Qwen3.5 4B, right? Where can I find that?
In order to combine Qwen3 0.6B with Qwen3.5 4B, the original Anima 3.8B uses custom nodes, but they only work with the original Anima 3.8B and no other models. So both Liquidn2 (Anima 3.8B LiquidMix) and bestluner180 (Anima Delta Mix 3.8B Qwen 3.5) created their own additional custom nodes to attach their models properly. As far as I understand, this is necessary to properly use Qwen3.5 4B with Anima. So if you cannot provide something like this as well, I don't it is even possible to use your model with Qwen3.5 4B...
@Lizardon1025
In the case of Forge Neo
put into Models\TextEncoders
https://huggingface.co/lylogummy/Anima-3.8B/tree/main/text_encoders
I'm not entirely sure if it's necessary, since I can generate images using the combination of qwen_image_vae and qwen_3_0.6_base.
@PPP_sleepy Yes, that sadly sounds like it only uses Qwen 3 0.6B...
@Lizardon1025
Unfortunately, I didn't merge it because I understood how it worked—it was just a fluke.
Time will take care of it eventually.
@Lizardon1025 @PPP_sleepy Quick note on the Qwen3.5 path, since it came up.
Any 3.8B checkpoint can run on native Qwen 0.6B. That is the chassis CLIP, so a merge usually still loads.
Qwen3.5 4B is different. On 3.8B the Gen-2 inserts include adaLN in four blocks. Those keys are tied to the 4B connector path. If a bake or injector overwrites them, the file still opens, but Q35 sampling goes black. That is a write problem, not a missing workflow screenshot.
Official Anima 3.8B nodes only attach the official checkpoint. Third-party 3.8B merges need their own attach / write path for those G2 slots. Delta Mix does that in ComfyUI; it does not repair a dead bake inside Forge Neo by itself.
So if this file is black on Q35, it is very likely the injector wrote over the G2 adaLN. Native 0.6B can still work. Q35 would need a clean attach, not the baked keys as they are.