Thank you for this! I am working on Anima based on ComfyUI code, but the GPL 3.0 license is a problem. This is based on ASL 2.0, which is great.

Please note that there is a possibility that some changes will be made after merging to align with other models.

If you need any help, feel free to reach out. I’d be happy to help.

I have a few questions.

1. To reduce repository size, the tokenizer settings are not stored in the repository but downloaded from HuggingFace. I'd like to use the official tokenizer config from https://huggingface.co/Qwen/Qwen3-0.6B. However, the official config seems to have a different value for max_position_embeddings (40960) than the settings in this PR (and ComfyUI) 32768: https://github.com/Comfy-Org/ComfyUI/blob/17e7df43d19bde49efa46a32b89f5153b9cb0ded/comfy/text_encoders/llama.py#L92. Can we use the official tokenizer config?
2. Also, regarding T5XXL, is it okay to use the official tokenizer config?: https://huggingface.co/google/t5-v1_1-xxl The tokenizer config in this PR seem to be slightly different from both the official and ComfyUI settings.

Yes, I noticed these issues as well when I reviewed them. However, these configurations come from the diffusion-pipe repository, and the owner of this repository is also the creator of this model.

Yes, I noticed these issues as well when I reviewed them. However, these configurations come from the diffusion-pipe repository, and the owner of this repository is also the creator of this model.

Thank you for the explanation. I understood. So keeping the config in the repository sounds good.

I also noticed one other thing: ComfyUI's inference code doesn't seem to use an attention mask in its Text Encoder calls, but diffusion-pipe does seem to use one.

I’m also very confused about this, but when I tested the model with and without an attention mask, I observed that using the attention mask makes the model more stable and produces better results.

Hmm, the attention mask is used for Text Encoder, but then the embedding where mask=0 is set to zero, and a fixed-length embedding is passed to DiT. That's interesting...

I'm not entirely sure about the author's intention, but here is my understanding (I could be wrong):

• Qwen3 needs mask internally , tokens self-attend, so padding must be excluded.
• After Qwen3, padding embeddings are zeroed out. In DiT cross-attention, K=0 produces near-zero scores, and V=0 contributes nothing. So zeros act as a "soft mask" automatically.

Thank you for the detailed explanation. I think I have a general understanding of Anima's training and inference.

Please understand that I will make some changes after merging. I would also appreciate any feedback or additional pull requests after the merge.

Thank you again for this great PR!

working well. Thank you @kohya-ss, @duongve13112002