v-prediction is Fast! Vivid! But Doesn’t Blend Well?
- v-pred is fast
- Colors are vivid
- Expressions are stiff
Introduction
Hello, I’m Easygoing.
This time, we’ll take a clear and simple look at v-prediction, a term you might occasionally hear about in the context of Stable Diffusion.
Starting with e-pred!
The core mechanism of modern image generation AI is called the diffusion model, which generates images by removing noise from a noisy starting point.
gantt
title e-pred and v-pred
dateFormat YYYY-MM-DD
tickInterval 6month
section e-pred
Stable Diffusion 1 :done, a1, 2022-08-22, 2025-05-01
section v-pred
Stable Diffusion 2.0 : b1, 2022-11-24, 2025-05-01
section e-pred + v-pred
Stable Diffusion XL 1.0 :done, c2, 2023-07-27, 2025-05-01
section Flow-matching
Stable Diffusion 3 : d1, 2024-06-12, 2025-05-01
AuraFlow : d2, 2024-07-12, 2025-05-01
Flux.1 : d3, 2024-08-01, 2025-05-01
HiDream-I1 : d4, 2025-04-06, 2025-05-01
The first practical open-source image generation AI was Stable Diffusion 1.
In Stable Diffusion 1, a standard method called e-pred (noise prediction) was used to remove noise.
Thinking of Noise and Images Separately!
In February 2022, a new approach was proposed to speed up image generation by separating noisy images into clean images and noise, then removing more of the noise component.
flowchart LR
A1(Noisy Image)
subgraph e-pred
B1(Capture Noisy Image)
B2(Remove Noise)
end
subgraph v-pred
C1(Capture as Clean Image + Noise)
C2(Efficiently Remove Noise Component)
end
D1(Next Image)
A1-->B1
B1-->B2
B2-->D1
A1-->C1
C1-->C2
C2-->D1
- e-pred (epsilon-prediction): Noise prediction
- v-pred (velocity-prediction): Velocity prediction
This method learns the “flow velocity” from noise to a clean image, earning the name v-pred (velocity-prediction).
v-pred converges faster, allowing images to be generated in roughly half the number of steps compared to traditional methods.
v-pred Models Produce Vivid Colors
When AI learns images, dark colors have low RGB values, making them harder to distinguish from noise.
As a result, traditional e-pred models struggled with poor rendering of dark colors and high-contrast expressions.
In contrast, v-pred models theoretically learn from less noisy images, making them better at rendering dark colors compared to e-pred models.
v-pred models enable higher contrast and more vivid color rendering than e-pred models.
v-pred Models Don’t Blend Well!
Image generation AI can achieve various expressions by controlling noise.
Example 1: Increasing detail with noise methods
Example 2: Advanced noise control with Detail Daemon
e-pred models can blend images by adding noise, making them adaptable to various images.
However, v-pred models separate noise and clean images even when noise is added midway, making them less likely to blend.
v-pred models struggle with expressions that require model switching, and when used with high-resolution techniques like Hires.fix or Detailer, they tend to produce stiff expressions that don’t blend well with the original image.
SDXL Returns to e-pred!
Let’s revisit the first chart.
gantt
title e-pred and v-pred
dateFormat YYYY-MM-DD
tickInterval 6month
section e-pred
Stable Diffusion 1 :done, a1, 2022-08-22, 2025-05-01
section v-pred
Stable Diffusion 2.0 : b1, 2022-11-24, 2025-05-01
section e-pred + v-pred
Stable Diffusion XL 1.0 :done, c2, 2023-07-27, 2025-05-01
section Flow-matching
Stable Diffusion 3 : d1, 2024-06-12, 2025-05-01
AuraFlow : d2, 2024-07-12, 2025-05-01
Flux.1 : d3, 2024-08-01, 2025-05-01
HiDream-I1 : d4, 2025-04-06, 2025-05-01
- Stable Diffusion 1: e-pred
- Stable Diffusion 2: v-pred
- Stable Diffusion XL: e-pred (+ v-pred)
The first Stable Diffusion 1 used the standard e-pred model.
Stable Diffusion 2, released in November 2022, adopted v-pred, but it wasn’t widely adopted due to a lack of compatibility with the well-established Stable Diffusion 1 community.
Stable Diffusion XL, released in July 2023, returned to e-pred, addressing e-pred’s weaknesses in detail and contrast by using a dedicated Refiner model.
However, the Refiner was difficult to fine-tune and slowed down illustration generation, so it didn’t gain widespread use. Instead, adjusting colors using the base model with CFG scale or extensions became the mainstream approach.
Back to v-prediction!
The vivid colors and stiff expressions of v-pred models aren’t a significant issue for anime-style illustrations.
In December 2024, the NoobAI-XL series, trained on a total of 13 million high-quality illustrations, introduced v-pred models, bringing them back into use with SDXL.
Today, models that merge e-pred and v-pred have emerged (though merging is theoretically challenging, thanks to the trial-and-error efforts of model creators), improving usability.
Comparing e-pred and v-pred!
Now, let’s compare illustrations generated by e-pred and v-pred models.
flowchart TB
subgraph e-pred
A1(SDXL_Base)
B1(Animagine-XL 3.1<br>2024.3.21)
B2(Illustrious-XL_v0.1<br>2024.9.25)
B3(Illustrious-XL_v1.1<br>2025.2.18)
B4(NoobAI-XL<br>2024.10.8)
C1([blue_pencil-XL_v7.0.0<br>2024.6.23])
C2([anima_pencil-XL_v5.0.0<br>2024.6.25])
C3([illustrious_pencil-XL_v4.0.0<br>2025.3.29])
end
subgraph v-pred
B5(NoobAI-XL_V-pred-1.0<br>2024.12.22)
C4([noob_v_pencil-XL-v2.0.1<br>2025.4.7])
end
A1-->B1
A1-->C1
A1---->B2
B1-->C2
B2--->B3
B3-->C3
B2-->B4
B4---->B5
C1-->C2
C2-->C3
B5-->C4
C3-->C4
Models for Comparison
- anima_pencil-XL-v5.0.0 (e-pred)
- noob_v_pencil-XL-v2.0.1 (v-pred)
We’ll input the same prompt into each model and compare the generated illustrations.
Illustration 1: Space Travel
Illustration 2: Agent
Both models are merged from blue_pencil-XL, but the generated illustrations differ significantly.
At a glance, the color and detail rendering of the v-pred model on the right is superior.
However, soft expressions, such as character facial expressions, are better in the e-pred model on the left. A common drawback of v-pred is a tendency to shift toward warm tones, resulting in color burn.
Mixing It All Together!
Let’s design a workflow that combines the strengths of both models.
flowchart LR
subgraph noob_v_pencil-XL
A1(Sketch)
end
subgraph Refiner
B1(Break Down Once)
end
subgraph anima_pencil-XL
C1(Redraw and Finalize)
end
A1-->B1
B1-->C1
First, create a sketch using the vivid colors of the v-pred model.
Next, use the Refiner to break down the illustration, enhancing saturation and detail.
Finally, regenerate the image with the e-pred model to soften the overall look and naturally refine character expressions.
This creates a soft, colorful illustration distinct from simply increasing the CFG scale with a v-pred model!
Conclusion: Try Using v-pred!
- v-pred is fast
- Colors are vivid
- Expressions are stiff
v-pred models offer vivid colors and details that traditional e-pred models couldn’t achieve.
v-pred is particularly suited for anime models, and development is progressing in the hugely popular illustrious-XL series.
However, v-pred has its weaknesses, and e-pred excels in diversity and soft expressions.
Moving forward, I hope to combine the strengths of both to explore unique expressions.
Thank you for reading to the end!