[Paper Quick Read] High-Quality Text-to-Image Generation in One Step—The Key to Extending MeanFlow from Class Labels to Text Was "Discriminability and Disentanglement of Text Encoders"

Hello! This is AI Nest, Inc.

You can also read explanations of related papers here.

Image generation with diffusion models requires dozens of sequential denoising steps, and the high inference cost has been a bottleneck for practical application. A promising direction to solve this is "Few-step Generation," and the ideal is generating high-quality images in a single step. MeanFlow (Geng+, 2025) is an approach that is realizing this dream through a beautiful formulation of "predicting average velocity," and it has achieved results on par with existing methods in class-label conditional generation on ImageNet.

However—there has been almost no research extending MeanFlow to text conditioning (T2I). Extending from class labels to text is a qualitatively different problem, moving from "1,000 ImageNet classes to infinite natural language," and simply applying existing T2I training losses does not work well.

The paper published this time by a team from Nankai University and Alibaba AMAP is one that has broken through this wall head-on. They successfully extended MeanFlow to T2I generation for the first time, achieving quality in 4 steps comparable to 30 steps of BLIP3o-NEXT. The key was identifying two properties that a text encoder should possess—Discriminability and Disentanglement. Let's read through this paper.

Figure 1 shows a comparison of 4-step outputs between the proposed method (Ours = EMF) and the current strongest distilled 1-step model, "SANA-Sprint." For complex prompts such as "a small city inside a vintage lightbulb," "a portrait of a conductor made of ink and smoke," "a high-altitude balloon overlooking a volcanic eruption," "a violin carved from ice," "a samurai in autumn leaf armor," and "a greenhouse on the moon," EMF clearly outperforms in detail accuracy and semantic fidelity. The words in blue are instruction elements that SANA-Sprint failed to capture.

What is MeanFlow—The innovation of predicting "average velocity"

Before we dive into the main topic, let's cover the basics of MeanFlow (Geng+, 2025).

Standard diffusion models and Flow Matching generate images by predicting the "instantaneous velocity at each time" and integrating the ODE. This is accurate, but requires many steps (usually 20-50) for integration. MeanFlow reverses this concept, enabling generation in principle in a single step by directly predicting the "average velocity" from time t to r.

Specifically, the flow map uθ(zt, t, r) directly predicts the "transition from zt to zr." In other words, the relationship is zr = zt + (r - t)・uθ(zt, t, r). To learn this, they derive a training objective function that connects the relationship between average velocity and instantaneous velocity via differential equations, which can be efficiently calculated using JVP (Jacobian-Vector Product).

With this idea, MeanFlow achieves few-step generation in principle without using heuristics like Consistency Models. It has already achieved competitive results in FID for class-label conditional generation (such as ImageNet 256x256), and many follow-up studies have emerged. However, extension to text conditioning remained untouched.

Why are class labels and text so different?

Extending from class labels to text is a more difficult problem than it appears.

Class labels have discrete and separated embedding spaces. "Dog," "cat," and "car" are sparsely distributed in the embedding space, maintaining large margins from each other. From the perspective of MeanFlow, this means that the "denoising trajectory becomes smooth." Since the instantaneous velocity at each time is close to and aligned with the average velocity, predicting the average velocity is easy. This is the structural reason why stable results can be obtained even with few-step generation.

On the other hand, natural language text embeddings are dense and continuous. Semantically similar expressions (e.g., "blue teapot" and "red teapot") are adjacent in the embedding space, and it is necessary to distinguish fine attribute differences. This causes the denoising trajectory to become a winding path, increasing the divergence between instantaneous velocity and average velocity. The authors call this "semantic drift," which becomes a fatal problem in few-step generation because there are limited opportunities for correction.

In addition, natural language often involves "prompts containing multiple attributes, objects, and spatial relationships" (e.g., "a blue ceramic teapot containing red tulips is placed on a wooden table"). To process such complex instructions in a few steps, the text representation itself must be able to "separate and hold sub-elements."

Two key properties: Discriminability and Disentanglement

After careful analysis, the authors identified two essential conditions for a text encoder to make MeanFlow-based T2I generation successful.

The first is Discriminability. This is the property where 'different prompts can generate distinguishable embeddings that are properly aligned with their corresponding images.' The verification method is cross-modal retrieval: calculate prompt embeddings for the text encoder using 118k pairs from COCO 2017, and obtain the top cosine similarities with image embeddings. The retrieved images are encoded with DINOv3 to measure similarity with the original query image. The results are BLIP3o-NEXT 0.734 > CLIP 0.730 > Gemma 0.713 > T5 0.634.T5 (a pure language model) is the lowest, while BLIP3o-NEXT and CLIP, which use image-text contrastive learning, rank higher.

The second is Disentanglement. This means the property where 'the embedding distance between a full prompt and its subsequences is small, and the linear structure is maintained.' Using all prompts from DPG-Bench, the cosine distance between the embeddings of the full version and a version with parts randomly removed was measured. The scores are BLIP3o-NEXT 0.999 > Gemma 0.987 > CLIP 0.967 > T5 0.893. LLM-based text encoders have an advantage, and BLIP3o-NEXT, which underwent image-text contrastive learning, is the clear winner.

By achieving both of these properties, MeanFlow can learn a stable average velocity even in few steps and faithfully render complex prompts into images.

EMF Design: MeanFlow Extension Based on BLIP3o-NEXT

Based on these insights, the authors propose EMF (Extending MeanFlow to T2I).

BLIP3o-NEXT (3B, 3B+) is adopted as the base model. To handle the two time conditions required by MeanFlow (interval length t-r and interval end t), the time embedding layer is duplicated into two, and provided to the flow map in the form φcond(t,r) = φinterval(t-r) + φend(t). The design constructs the velocity network uθ(zt, t, r, ψtext) in combination with the text embedding ψtext.

The training data consists of approximately 170,000 samples (BLIP3o-60k, shareGPT-4o, Echo-4o). The learning rate is 1e-5, batch size is 128, and it runs for 150 epochs. There is a trick to sampling training times, adopting adaptive sampling with a Beta distribution to 'learn both short-term and long-term intervals in a balanced manner.'

The training objective function is close to the standard MeanFlow loss, in the form of ||uθ(zt, t, r, ψtext) - sg(ũtgt)||², but the target ũtgt includes 'instantaneous velocity + interval error term' and is stabilized by Stop-Gradient.

Experimental Result 1: GenEval shows 4-step performance on par with 30-step BLIP3o-NEXT

Now, let's look at the experimental results. GenEval is a rigorous metric focused on attributes.

Figure 4 shows the progression of GenEval scores against training steps (in W units, 1W=10k steps). EMF improves rapidly in 1-step, 2-step, and 4-step, reaching a 4-step GenEval of 0.90 at approximately 60k steps. This is a value almost comparable to the 30-step version of BLIP3o-NEXT (0.91). Even at 2-step (0.85) and 1-step (0.74), it reaches a practical range even with extremely few steps.

What is noteworthy is the property that 'quality improves as the number of steps increases.' While it is known that consistency-based models can suffer from quality degradation as steps and expense increase, EMF stably improves from 2-step to 4-step, and further to 8-step (DPG-Bench 81.20→81.94). This is evidence that the essential characteristic of MeanFlow, which is 'learning average velocity,' is maintained even under text conditions.

Experimental Result 2: Surpassing SOTA across the board on DPG-Bench / HPS-v2.1

The results on DPG-Bench, which evaluates more complex compositional prompts, and HPS-v2.1, which measures human preference, are also dramatic.

In 1-step DPG-Bench, EMF recorded 77.36, an improvement of +20.31 points over the 1-step BLIP3o-NEXT (57.05). It also showed 79.44 (+12.06) at 2-step and 81.20 (+3.05) at 4-step; the improvement margin is larger with fewer steps, and at 4-step, it comes within a hair's breadth of the 30-step baseline (82.05).

HPS-v2.1 (human preference) is even clearer, with improvements of +7.23 at 1-step, +4.76 at 2-step, and +2.29 at 4-step. Consistent improvements were seen across all categories: Anime, Concept-Art, Paintings, and Photo, confirming a significant improvement in visual appeal as well.

Even in absolute values, the 3B parameter, 4-step EMF GenEval score of 0.90 surpasses major pre-trained models such as SD3.5-L (8B, 28-step, 0.71), FLUX.1-dev (12B, 50-step, 0.66), PixArt-α (0.6B, 20-step, 0.48), and SANA-1.5 (4.8B, 20-step, 0.81).

Comparison with existing distillation-based models: Outperforming all distillation models

Another mainstream approach for few-step generation is 'Distillation.' This is a method of distilling a few-step student model from a teacher model, and there are many such as SDXL-LCM, SDXL-Turbo, Hyper-SDXL, SANA-Sprint, and rCM.

Figure 5 is a qualitative comparison of 4-step generation. For prompts that are difficult for any existing distillation model—such as SANA-Sprint, Flux.1-Schnell, SDXL-DMD2, or SD3.5-L-Turbo—like 'A long-exposure glow of a mangrove in moonlight where the roots form musical notation and bioluminescent notes drift on the tide' or 'A cyborg falcon scattering Mach circle vapor in mid-air, compressed by a super-telephoto lens,' EMF is clearly superior in both semantic fidelity and visual detail. Blue text indicates elements where other models failed.

In quantitative comparison, EMF (3B, 4-step, GenEval=0.90) outperformed the strongest distillation model, rCM (14B, 4-step, 0.83), by 7 points with 4.6 times fewer parameters. It also clearly distinguished itself from representative distillation models such as SANA-Sprint (1.6B, 4-step, 0.77) and SDXL-DMD2 (2.6B, 4-step, 0.58).

Inference time: 15.5x speedup from 1.24 seconds to 0.08 seconds

The most important aspect in practical terms is inference speed.

When generating the same prompt on an H200 GPU, the 30-step version of BLIP3o-NEXT takes 1.24 seconds. EMF takes 0.22 seconds at 4-step, 0.12 seconds at 2-step, and 0.08 seconds at 1-step, achieving a maximum speedup of 15.5x. With a 15.5x speed improvement at 1-step and a 5.6x improvement even at 4-step, it reaches the realm of near real-time response while maintaining quality.

This has decisive implications for API services and edge deployment. If one image can be generated in under 100ms even with complex prompts, applications for interactive image editing UIs and video generation will expand significantly.

Ablation: Why does SANA-1.5 fail?

The reverse verification is also interesting. The authors attempted to train EMF with SANA-1.5 (a Gemma-based text encoder similar to CLIP), but failed.

The cause is the lack of discriminative power in the text encoder. Even when the SANA-1.5 encoder is fine-tuned with SFT data, while the quality of Flow Matching (multi-step) improves from 0.81 to 0.85, MeanFlow training itself still fails to converge, remaining at GenEval=0.47 at 4-step. This shows that it is not the case that 'any powerful LLM-based encoder will do,' but rather that only encoders that have acquired high discriminative power and discreteness through image-text contrastive learning function in MeanFlow-based T2I.

A similar trend was confirmed with the OpenUni text encoder (InternVL3-based Gemma), where the result was that while Gemma has high discreteness, it lacks sufficient discriminative power, leading to failure in MeanFlow-based few-step generation.

Summary: The first step to pushing MeanFlow into practical T2I generation

What EMF has achieved is bringing the theoretically beautiful framework of MeanFlow into practical T2I generation for the first time.

・First T2I extension of MeanFlow: The first study to succeed in natural language conditional generation using MeanFlow, which was previously limited to class labels.

・Discovery of discriminative power and discreteness: Identified two essential conditions for text encoders to succeed in few-step generation.

・SOTA-class efficiency and quality: Outperforming larger pre-trained models and distillation models across the board with 3B parameters, 4-step, and GenEval=0.90.

・15.5x speedup: BLIP3o-NEXT 30-step (1.24 seconds) → EMF 1-step (0.08 seconds).

The authors have released the code (github.com/AMAP-ML/EMF), and high reproducibility is also a feature. I feel this paper is an excellent answer to the homework of 'elevating MeanFlow to practical T2I' and will serve as a guideline for future few-step generation research.

It is common knowledge in image generation that 'the choice of pre-trained model determines everything,' but this paper adds a new perspective to that common knowledge: 'No, the discriminative power and discreteness of the text encoder are the deciding factors in the era of few-step generation.' This is likely to become an insight that both designers and users of Foundation Models cannot afford to overlook.

For those who want to organize the papers they have read so they can easily look back on them later, or for those who want to share them with their team, the paper management service 'Scacon' is also recommended. It is also quite compatible with organizing and accumulating research for those who write papers.