UniMore-BAGEL

The final UniMore checkpoint based on BAGEL-7B-MoT, supporting multimodal understanding and image generation. This repository contains the merged BF16 model, its VAE, tokenizer, and configuration files.

Method

Two branches are trained independently from the same pretrained BAGEL model:

• Generation: grounded phrase-to-crop supervision with the native generation loss.
• Understanding: full-image caption supervision with positive cross-entropy and full-caption ranking against detail-omission, incorrect-detail, and random negatives.

The branches are combined using

W = W_base + ((W_t2i - W_base) + lambda * (W_i2t - W_base))

Here, lambda controls the understanding branch's contribution. Merge arithmetic uses BF16, including intermediate differences and additions.

Download and Use

hf download HuggingSelf/UniMore --local-dir /data/models/UniMore-BAGEL

Use the BAGEL adapter in TorchUMM, setting model_path to the downloaded directory. Installation, training, and inference instructions are in the UniMore package. These weights use BAGEL's native ema.safetensors format, not a Transformers AutoModel checkpoint.

The generation evaluation configuration uses 512px images, 50 sampling steps, text/image CFG 4/1, timestep shift 3, and seed 42.

Training

Setting	Generation	Understanding
Training records	2,460	4,000
Optimizer updates	1,000	600
Learning rate	5e-5	1.8e-5
Warmup updates	50	60
Training image size	256	112
LoRA rank / alpha	16 / 32	16 / 32

Training data and a preparation script are available in the UniMore dataset repository.

Files and Attribution

• ema.safetensors: final merged model.
• ae.safetensors: the original BAGEL VAE.
• Tokenizer and JSON files: the original BAGEL inference configuration.
• release.json: file sizes and fixed source revisions.

The base model is BAGEL-7B-MoT. The Apache 2.0 license is included in LICENSE.