Add NewbiePipeline and NextDiT_3B_GQA_patch2_Adaln_Refiner_WHIT_CLIP transformer

[E-Anlia]

This PR introduces a new text-to-image pipeline named NewbiePipeline, as well as a new
NextDiT-based transformer architecture,
NextDiT_3B_GQA_patch2_Adaln_Refiner_WHIT_CLIP, fully implemented following
Diffusers' pipeline and model design principles.

🚀 Main additions

• New pipeline
Adds NewbiePipeline under diffusers.pipelines.newbie/.
The pipeline follows the standard Diffusers structure (DiffusionPipeline subclass) and
supports loading via from_pretrained.

• New transformer architecture
Adds transformer_newbie.py, implementing:

• NextDiT backbone with grouped-query attention (GQA)
• Adaln-Refiner blocks
• Patch-size 2 vision encoder
• 36 transformer layers
• 2304 hidden dims
• WHIT CLIP–style text conditioning

The transformer inherits from ModelMixin, enabling standard save/load, weight
serialization and integration with Diffusers utilities.

• RMSNorm implementation
Adds RMSNorm to diffusers.models.components, using a PyTorch fallback and supporting
Apex fused RMSNorm if available.

• Scheduler compatibility
The pipeline is compatible with FlowMatchEulerDiscreteScheduler without requiring
additional custom scheduler code.

🧩 Motivation

This PR provides an implementation of a modern NextDiT-style text-to-image architecture
with high-resolution capability and strong conditioning support.
The goal is to enable researchers and users to load, run, and fine-tune this model
directly through Diffusers with minimal friction.

📁 Files added

src/diffusers/models/components.py
src/diffusers/models/transformers/transformer_newbie.py
src/diffusers/pipelines/newbie/pipeline_newbie.py
src/diffusers/pipelines/newbie/init.py

shell
Copy code

📁 Files modified

src/diffusers/init.py
src/diffusers/models/init.py
src/diffusers/models/transformers/init.py
src/diffusers/pipelines/init.py

yaml
Copy code

✔ Notes

• No external dependencies required
• Apex is optional; PyTorch RMSNorm is the default path
• The pipeline has been tested locally with from_pretrained and produces expected outputs
• Follows the established structure of Diffusers pipelines & transformer modules

Fixes # (no issue linked)

---

Before submitting

• I have read the contributor guidelines
• This PR introduces a new pipeline and model
• All necessary registration points are updated
• The implementation is consistent with existing Diffusers conventions

Who can review?

Tagging pipeline & transformer reviewers:
@asomoza @yiyixuxu @sayakpaul

[sayakpaul]

Can you link the original codebase, paper, and some results of this model?

[E-Anlia]

https://huggingface.co/NewBie-AI/NewBie-image-Exp0.1
https://github.com/[NewBieAI-Lab/NewBie-image-Exp0.1

This model is based on improvements made to research on lumina.
Based on NextDiT
Example：