Community: config-driven Lens-Base LoRA trainer (flow-match, Comfy export) — working on 16GB

[LoboForge]

Hi Lens team,

We've built and validated a standalone LoRA trainer for microsoft/Lens-Base and wanted to share it in case it's useful to the community or future official tooling.

Trainer repo: https://github.com/LoboForge/LoboForge-LensTrainer
Hugging Face Space: https://huggingface.co/spaces/LoboForge/LoboForge-LensTrainer
HF collection: https://huggingface.co/collections/LoboForge/lens-training-loboforge

What works today

• Config-driven training: python train.py configs/train_lora_lens_base_24gb.yaml
• Flow-match loss aligned with LensPipeline (latents [B, H×W, 128], GPT-OSS multi-layer text, timestep/1000)
• PEFT LoRA on LensTransformer2DModel; TE/VAE frozen
• ComfyUI-compatible export (diffusion_model.* key remap)
• 24GB preset: CPU offload, TE + latent cache, disable_mxfp4, grad checkpointing, AdamW 8-bit
• Checkpoint resume, mid-training samples, loss.json

Example result (community LoRA, not Microsoft-endorsed)

• Writeup: https://github.com/LoboForge/LoboForge-LensTrainer/blob/main/docs/loras/sebastian-jessica-v2.md
• Weights (HF): https://huggingface.co/LoboForge/lens-lora-sebastian-jessica-v2

Environment

• Lens installed from vendor/Lens (git clone; upstream has no pyproject.toml)
• Tested on consumer/datacenter GPUs with Lens-Base local or HF hub

We're not asking for an immediate merge — mainly flagging that Lens-Base LoRA training is practical with the public inference package, and happy to answer questions or share configs/logs if helpful.

Thanks for open-sourcing Lens!