qwen3-vl-flux2-8b

My Custom Qwen3-VL-8B

Fine-tuned on 20k image-flux.2 dev JSON pairs.

Usage

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
import torch

model = Qwen3VLForConditionalGeneration.from_pretrained("deepglugs/qwen3-vl-flux2-8b", torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained("deepglugs/qwen3-vl-flux2-8b")
# ... your inference code
# Load your image
image_path = "path/to/your/image.png"
image = Image.open(image_path).convert("RGB")

# Use the special flux.2 prompt
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "Describe this image in the flux.2 JSON format."}
        ]
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to("cuda")

# Generate (use temperature for slight variation if desired)
generated_ids = model.generate(
    **inputs,
    max_new_tokens=1024,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
)
generated_text = processor.batch_decode(generated_ids[:, inputs.input_ids.shape[-1]:], skip_special_tokens=True)[0]

# Parse and pretty-print the JSON output
try:
    json_output = json.loads(generated_text)
    print(json.dumps(json_output, indent=2))
except json.JSONDecodeError:
    print("Raw output (JSON parsing failed):")
    print(generated_text)