SciForma-9B

SciForma-9B is the final M-DPO aligned checkpoint of SciForma: Structure-Faithful Generation of Scientific Diagrams, fine-tuned from SciForma-Base via Multi-Dimensional Conjunctive DPO (M-DPO).

Model	Overall ↑	Comp. ↑	Arrow ↑	Text ↑
SciForma-9B	69.51	74.49	66.46	67.00
GPT-Image-1.5	68.96	75.50	62.50	68.20
SciForma-Base	67.59	73.52	64.64	63.84
FLUX.2-klein-base-9B (no fine-tuning)	33.87	42.80	25.20	23.60

Scores on SciFormaBench-2K (GPT-5.4 judge, split_dims=True).

Usage

This repository contains only the fine-tuned transformer weights. The text encoder, tokenizer, VAE, and scheduler are inherited from the base model.

import torch
from diffusers import Flux2KleinPipeline, Flux2Transformer2DModel

# Load fine-tuned transformer
transformer = Flux2Transformer2DModel.from_pretrained(
    "microsoft/SciForma-9B",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)

# Load full pipeline with base model components
pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9B",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()

# Generate
prompt = "The figure illustrates a transformer architecture with encoder and decoder..."
image = pipe(
    prompt,
    height=576,
    width=1008,
    num_inference_steps=28,
    guidance_scale=4.0,
).images[0]
image.save("output.png")

Training Details

• Base model: SciForma-Base (Stage 2 SFT)
• M-DPO: Multi-Dimensional Conjunctive DPO with 1v2+global-worst triples (~16.5K groups), targeting Component and Text axes simultaneously
• Hardware: 4× B200, batch=1/GPU, GA=3 → eff_batch=12, 10K steps
• Training data: microsoft/SciFormaData-700K

License

Released under the FLUX Non-Commercial License v2.1. Non-commercial research use only.

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Non-Commercial License. Copyright Black Forest Labs Inc. This model has been modified by Microsoft Research Asia.