Principle:Deepseek ai Janus ODE Denoising

Knowledge Sources	Flow Matching for Generative Modeling JanusFlow: Harmonizing Autoregression and Rectified Flow Janus: Decoupling Visual Encoding
Domains	Image_Generation, Diffusion_Models
Last Updated	2026-02-10 09:30 GMT

Overview

An iterative denoising procedure that solves an ODE to transport latent noise into a clean image representation, using an LLM as the velocity predictor with ShallowUViT encoder/decoder for latent-to-LLM bridging.

Description

The ODE denoising loop is the core generation mechanism in JanusFlow. Unlike autoregressive methods that generate tokens sequentially, rectified flow generates images by iteratively refining a noisy latent through Euler ODE steps. At each step:

1. Encode latent: ShallowUViTEncoder processes the current noisy latent with a timestep embedding
2. Align to LLM: Linear aligner projects UViT output (768-dim) to LLM dimension (2048-dim)
3. LLM forward: The language model processes the concatenated text + timestep + latent embeddings
4. Align from LLM: RMSNorm + linear aligner projects LLM output (2048-dim) back to UViT dimension (768-dim)
5. Decode velocity: ShallowUViTDecoder predicts the velocity field from the projected hidden states
6. CFG: Conditional and unconditional velocities are combined
7. Euler step: The latent is updated: z = z + dt × v

KV-caching is used to avoid recomputing prompt tokens after the first step.

Usage

Use this principle after noise initialization to denoise the latent over num_inference_steps (default 30) iterations.

Theoretical Basis

The rectified flow ODE:

$\frac{d{z}_t}{dt}=v_{\theta }(z_t,t,c)$

Solved with the Euler method:

$z_{t+dt}=z_t+dt\cdot v_{\theta }(z_t,t,c)$

Where v_θ is the velocity field predicted by the combined ShallowUViT-LLM pipeline, and c is the text conditioning.

CFG for velocity:

$v=w\cdot v_{cond}-(w-1)\cdot v_{uncond}$

The timestep is normalized: t = step / num_steps × 1000.

Related Pages

Implemented By

• Implementation:Deepseek_ai_Janus_ODE_Denoising_Loop

Uses Heuristic

• Heuristic:Deepseek_ai_Janus_Bfloat16_Operation_Workarounds