Implementation:Deepseek ai Janus Apply Sft Template JanusFlow

Knowledge Sources	Janus
Domains	NLP, Image_Generation
Last Updated	2026-02-10 09:30 GMT

Overview

Concrete tool for formatting conversation prompts using the JanusFlow VLChatProcessor's SFT template for rectified flow generation.

Description

The JanusFlow VLChatProcessor.apply_sft_template_for_multi_turn_prompts method formats conversations identically to the standard Janus processor. The image_start_tag ("<begin_of_image>") is appended to trigger generation. The JanusFlow processor also defines image_gen_tag ("<｜begin▁of▁generation｜>") at processing_vlm.py:L91.

Usage

Call this method before CFG input preparation in the JanusFlow generation pipeline.

Code Reference

Source Location

• Repository: Janus
• File: janus/janusflow/models/processing_vlm.py
• Lines: L159-199 (apply_sft_template_for_multi_turn_prompts), L89 (image_start_tag), L91 (image_gen_tag)

Signature

class VLChatProcessor(ProcessorMixin):
    def apply_sft_template_for_multi_turn_prompts(
        self,
        conversations: List[Dict[str, str]],
        sft_format: str = "deepseek",
        system_prompt: str = "",
    ) -> str:
        """Format conversations into SFT prompt string."""

    @property
    def image_start_tag(self) -> str:
        """Returns '<begin_of_image>'."""

    @property
    def image_gen_tag(self) -> str:
        """Returns '<｜begin▁of▁generation｜>'."""

Import

from janus.janusflow.models import VLChatProcessor

I/O Contract

Inputs

Name	Type	Required	Description
conversations	List[Dict[str, str]]	Yes	Message dicts with "role" and "content"
sft_format	str	No	Template name (default "deepseek")
system_prompt	str	No	System message (default "")

Outputs

Name	Type	Description
prompt	str	SFT-formatted prompt + image_start_tag, ready for tokenization

Usage Examples

JanusFlow Prompt Formatting

messages = [
    {'role': 'User', 'content': 'A beautiful mountain landscape at sunset'},
    {'role': 'Assistant', 'content': ''}
]

text = vl_chat_processor.apply_sft_template_for_multi_turn_prompts(
    conversations=messages,
    sft_format=vl_chat_processor.sft_format,
    system_prompt=''
)
text = text + vl_chat_processor.image_start_tag
input_ids = torch.LongTensor(tokenizer.encode(text))

Related Pages

Implements Principle

• Principle:Deepseek_ai_Janus_Prompt_Formatting_for_Flow_Generation

Uses Heuristic

• Heuristic:Deepseek_ai_Janus_Image_Generation_Prompt_Tips