[New Model]: circlestone-labs/Anima

[timzsu]

The model to consider.

https://huggingface.co/circlestone-labs/Anima

Anima is a 2B text-to-image diffusion model for anime, character, and illustration generation.

The closest model vllm-omni already supports.

Closest model: Qwen/Qwen-Image, supported by QwenImagePipeline.

Similarity: Anima is also a text-to-image diffusion model, and its released files include qwen_image_vae.safetensors, so part of the Qwen-Image VAE path may be reusable.

Difference: Anima is is packaged as a Diffusion Single File model, while Qwen/Qwen-Image is supported through vLLM-Omni's native Qwen-Image pipeline/repo layout.

What's your difficulty of supporting the model you want?

The main difficulty is Diffusion Single File support. vLLM-Omni currently supports repo-style diffusers loading through DiffusionPipeline.from_pretrained(), but does not support Diffusion Single File checkpoints directly.

Likely patch points:

• vllm_omni/diffusion/models/diffusers_adapter/pipeline_diffusers_adapter.py: DiffusersAdapterPipeline.load_weights() currently always calls DiffusionPipeline.from_pretrained(model_id, **load_kwargs). It needs a single-file path that calls DiffusionPipeline.from_single_file(...).
• vllm_omni/diffusion/data.py: OmniDiffusionConfig.enrich_config() currently expects diffusers repo metadata like model_index.json. Single-file support needs a way to skip or override that metadata path and provide the pipeline class/config explicitly.

Use case and motivation

Anima is a popular anime / illustration-focused text-to-image model. Supporting it would expand vLLM-Omni's image generation coverage and allow serving Anima through the existing image generation or diffusion chat APIs.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[timzsu]

@yuanheng-zhao

[akshatvishu]

I can work on this if it's still open!

[timzsu]

I can work on this if it's still open!

Hi @akshatvishu, feel free to work on this if you are interested.

[patelchaitany]

@akshatvishu, are you working on this issue if not then love to pick up otherwise happy to collaborate as well

[akshatvishu]

@akshatvishu, are you working on this issue if not then love to pick up otherwise happy to collaborate as well

It's under work; will link the pr in a short while!

[timzsu]

Hi @akshatvishu, it's been over two weeks since you claimed the task. How's the progress? Is there any difficulty when adding the model?

[akshatvishu]

Hi @akshatvishu, it's been over two weeks since you claimed the task. How's the progress? Is there any difficulty when adding the model?

Sorry for the delay and thanks for checking in! Due to a recent rebase of the vLLM version, my already-open PRs are taking longer than usual to get merged. I’ve pushed the initial WIP code and will benchmark it, then follow up here within the next 24 hours.